Equipment predictive maintenance system based on multi-modal data fusion and edge collaborative decision

Abstract

An equipment predictive maintenance system based on multi-modal data fusion and edge collaborative decision comprises intelligent data acquisition cards and a unified management platform, and the intelligent data acquisition cards are connected with each other through internal LoRa networking units to form a dynamic Mesh network; the host intelligent data acquisition card is in communication connection with the cloud through the network unit; the unified management platform comprises a cloud server and a user terminal, provides an AI model management function, and supports remote issuing of a trained MSAD algorithm model file to a specified intelligent data acquisition card or equipment group in a differential upgrade package form; the intelligent data acquisition card performs real-time analysis on the acquired multi-modal sensing data based on a built-in MSAD algorithm model, and outputs a quantized abnormal confidence score; and when the abnormal confidence score is within a preset uncertainty interval of the system, triggering a collaborative verification mechanism. According to the invention, the accuracy and reliability of equipment fault early warning are obviously improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and specifically relates to a predictive maintenance system and method for equipment based on multimodal data fusion and edge collaborative decision-making. Background Technology

[0002] Traditional data acquisition schemes have significant limitations: (1) Single function: Traditional acquisition cards only have analog-to-digital conversion and IO control functions. Data analysis is entirely dependent on the host computer and cannot perform real-time intelligent analysis at the edge of data generation. (2) Lack of collaboration: The data dimensions of single-point acquisition are limited and cannot be integrated with multimodal data (such as vibration, temperature, and current) across locations and types for comprehensive decision-making. Moreover, the acquisition points are isolated from each other and cannot form a collaborative sensing and control network. (3) Poor flexibility: The acquisition logic, control strategy, and AI model are all fixed by the manufacturer or rely on complex host computer software for secondary development, making it difficult to dynamically and adaptively adjust according to changes in on-site working conditions.

[0003] Existing data acquisition and edge analysis solutions have inherent defects: (1) Isolated and singular edge AI models: Existing edge computing solutions mostly target single data types or simple rules, and cannot effectively integrate cross-modal complementary information. They have limited ability to identify complex fault modes, and the model output lacks quantification of its own judgment uncertainty (such as confidence level), resulting in a crude decision-making mechanism. (2) Static and rigid networking methods: Existing distributed sensor networks mostly adopt preset master-slave nodes or static topologies, and cannot dynamically optimize the network structure and task allocation according to the resource status of on-site nodes (computing power, power, link quality), resulting in low overall system robustness and resource utilization efficiency. (3) Lack of collaborative decision-making mechanisms: When edge nodes encounter ambiguous abnormal states, existing solutions either directly report to the cloud (introducing delay) or rely on the uncertain judgment of a single node to make key decisions, lacking the ability to use neighboring nodes on the edge side for fast, low-latency collaborative verification and joint decision-making. Therefore, existing technologies cannot achieve highly reliable, adaptive, and resource-optimized intelligent operation and maintenance in industrial sites. This invention addresses the aforementioned shortcomings by proposing an integrated solution that enables multimodal data fusion, intelligent networking, and collaborative decision-making at the edge. Summary of the Invention

[0004] Purpose of the Invention: To overcome the above shortcomings, the purpose of this invention is to provide a predictive maintenance system for equipment based on multimodal data fusion and edge collaborative decision-making. This system effectively solves the problems of insufficient reliability of single-point decision-making, low utilization of static network resources, and a simplistic decision-making mechanism in traditional edge solutions. While maintaining the low latency advantage of the edge, it significantly improves the accuracy of fault warning and system reliability through intelligent collaboration between nodes. This invention addresses the lack of intelligent collaboration and ambiguous decision-making mechanisms in current data acquisition systems, significantly improving the accuracy and reliability of fault warning for industrial equipment.

[0005] Technical Solution: To achieve the above objectives, this invention provides a predictive maintenance system for devices based on multimodal data fusion and edge collaborative decision-making, comprising: an edge side, wherein the edge side includes at least two intelligent data acquisition cards, the intelligent data acquisition cards are deployed on the device for data acquisition and control, and the intelligent data acquisition cards are interconnected through an internal LoRa networking unit to form a dynamic Mesh network, one of the intelligent data acquisition cards is elected as the host node, and the others are collaborative nodes; the intelligent data acquisition card corresponding to the host node is connected to the cloud through a network unit; A unified management platform, comprising a cloud server and user terminals, is provided. The cloud server interacts with host nodes, and user terminals access the cloud server via the internet for monitoring and management. It also offers AI model management functionality, supporting the remote distribution of trained MSAD algorithm model files as differential upgrade packages to designated intelligent data acquisition cards or device groups. The specific differential upgrade process is as follows: 1) Differential Package Generation: The cloud server compares the new version of the MSAD algorithm model file with the old version of the model file byte by byte, extracts the difference (Delta) to generate a differential upgrade package, and calculates its MD5 value.

[0006] 2) Remote distribution: The cloud server splits the differential packet into data blocks and distributes them one by one to the designated target intelligent data acquisition card through the host node.

[0007] 3) Verification and Deployment: After the target acquisition card receives the data, it calculates the MD5 value of the received data and compares it with the original MD5 value sent from the cloud to ensure data integrity. After successful verification, the acquisition card uses its built-in bootloader program to merge the differential packet with the local old model, generate a new model, and load and run it.

[0008] 4) Safe rollback: If the new model fails to load or causes system instability, the acquisition card will automatically switch back to the old version and report the failure log to the management platform.

[0009] The intelligent data acquisition card performs real-time analysis on locally acquired multimodal sensor data based on the built-in MSAD algorithm model and outputs a quantified anomaly confidence score. When the anomaly confidence score is within the uncertainty range between the first high threshold and the second low threshold preset by the system, the collaborative verification mechanism is triggered.

[0010] The collaborative verification mechanism is jointly completed by the requesting node, the dynamically elected host node, and one or more collaborative nodes. It synchronously collects data, performs distributed reasoning and merges results, and finally forms a joint decision at the edge.

[0011] The intelligent data acquisition card described in this invention includes a data acquisition unit, a control unit, a core processing unit, a communication unit, and an auxiliary unit, all of which are bidirectionally connected to the core processing unit. The data acquisition unit includes an analog acquisition unit and a digital acquisition unit, and the control unit includes an analog control unit and a digital control unit; The core processing unit is based on the analysis and calculation unit, and the MSAD algorithm model is located within the analysis and calculation unit. The communication unit includes a networking unit and a network unit; the collaboration between the analysis and computing unit and the networking unit jointly supports the MSAD algorithm model and the dynamic collaboration mechanism. The auxiliary unit includes a power supply unit and an indicator unit.

[0012] The analysis and computing unit described in this invention adopts a heterogeneous computing architecture, which includes a multi-core CPU and a dedicated AI acceleration chip (NPU / FPGA) for running the multimodal perception anomaly detection MSAD algorithm model.

[0013] Furthermore, the networking unit is used to realize inter-node networking communication and uplink data transmission to the master node. The communication methods used for inter-node networking communication include, but are not limited to, LoRa wireless communication modules, and can also use Wi-Fi Mesh, Zigbee, Bluetooth Mesh or other communication technologies that support self-organizing networks; the uplink data transmission methods to the master node include, but are not limited to, wired Ethernet, wireless Wi-Fi, and cellular networks.

[0014] The MSAD algorithm model described in this invention adopts a dual-branch deep neural network structure, including a numerical anomaly perception branch for capturing local feature patterns and a sequence anomaly perception branch for capturing long-term sequence dependencies. The numerical anomaly perception branch adopts a one-dimensional convolutional neural network (CNN), and the sequence anomaly perception branch adopts a gated recurrent unit (GRU). The MSAD algorithm model performs fusion analysis on the collected multimodal time series data, outputs anomaly confidence scores and fault type labels, and triggers different levels of response according to preset multi-level threshold rules.

[0015] More preferably, the multi-level threshold rule specifically includes: If the abnormal confidence score is greater than the first high threshold, local control will be triggered immediately and a high-risk alarm will be reported. If the abnormal confidence score falls within the "uncertainty interval" between the first high threshold and the second low threshold, the collaborative verification mechanism is activated to request verification from neighboring nodes. If the anomaly confidence score is lower than the second lowest threshold, only the data is recorded and no active response is triggered.

[0016] Multi-level threshold calibration methods require calibration using historical data and are not fixed. The system provides a threshold self-learning function. 1): Collect historical data under normal equipment conditions and input it into the MSAD model to obtain the confidence distribution. Set the 95th percentile as the low threshold T_low. 2): Collect known fault injection or historical fault data to obtain the confidence distribution, and set the 5th percentile as the high threshold T_high. The examples given in this specification represent calibration results under specific test conditions only. For different equipment (such as fans, machine tools, pumps), the above calibration process needs to be re-executed in actual applications to adapt to different operating conditions.

[0017] The predictive maintenance system for equipment based on multimodal data fusion and edge collaborative decision-making described in this invention has the following data processing flow for the MSAD algorithm model: 1) Data preprocessing and alignment: After receiving the raw data collected by the data acquisition unit, a joint denoising algorithm based on wavelet transform is used to filter high-noise industrial data. The 'db4' wavelet basis function is used to perform a 5-level decomposition of the original signal, and a soft thresholding function is used to denoise the high-frequency detail coefficients. After denoising, the Dynamic Time Warping (DTW) algorithm is used to align the time-series asynchronous data between multimodal sensor data. During alignment, the curved window size is set to 10, and the distance metric is Euclidean distance to ensure the synchronization of different modal data (such as vibration and current) on the time axis. 2): Multi-scale feature extraction, which utilizes the parallel computing capabilities of the AI ​​acceleration chip in the analysis and computing unit. The pre-processed data will be simultaneously subjected to multi-scale feature value extraction in the time domain, frequency domain, and time-frequency domain. The system will then fuse the extracted feature values ​​into a single feature vector. 3): Anomaly detection and inference: The fused single feature vector is input into the MSAD algorithm model for inference. Specifically, a one-dimensional convolutional neural network of the numerical anomaly perception branch is used to capture local feature patterns, and a gated recurrent unit (GRU) of the sequence anomaly perception branch is used to capture long-term sequence dependencies. The outputs of the two branches, numerical anomaly perception and sequence anomaly perception, are aggregated in the decision fusion layer to form a fused comprehensive feature. The comprehensive feature is fed into the fully connected layer to finally generate a comprehensive anomaly confidence score (0-1) and a preliminary fault type label. This anomaly confidence score is the direct basis for triggering different levels of response (including collaborative verification). 4): Adaptive collaborative decision-making mechanism: The system performs different operations through adaptive decision-making based on preset multi-level threshold rules.

[0018] Preferably, in this invention, the intelligent data acquisition card uses a resource-aware distributed host election algorithm to elect a host node. During the host node election process, the resource status of each node is exchanged through heartbeat packets, and the host node is elected based on a dynamic scoring function. The host node then coordinates member nodes to perform data sharing and distributed collaborative computing, as detailed below: 1) Initial Network Setup and Status Broadcast: After powering on, each node enters network maintenance mode and broadcasts a "heartbeat" packet containing its own device ID, computing power (CPU utilization, memory remaining), power status, and signal strength (RSSI) every T seconds (T is set to 30 seconds by default, and the period can be configured remotely through the management platform). 2) Dynamic Host Node Election: After a node receives a heartbeat packet from another node, it does not immediately trigger a re-election to avoid network jitter. Each node maintains an election timer (the default duration is 3 times the heartbeat period, i.e., 90 seconds). When the timer expires, the node collects the latest heartbeat information from all neighboring nodes received within that window, substitutes it into the scoring function below to calculate a score, and the node with the highest score is identified as the host node. If no host node is specified, all nodes follow this rule, and the node with the highest score becomes the host node; other nodes automatically join the network as members. Because CPU_Free (percentage), Memory_Free (MB), Battery_Level (percentage), and RSSI (dBm, usually negative) have different units, they need to be Min-Max normalized before being substituted into the scoring function to uniformly map them to the [0, 1] interval. Specific formula: in, Let the minimum and maximum values ​​of this physical quantity be defined within the network during the most recent time window (where each node maintains a sliding window of fixed length L (e.g., L=10) to store the raw values ​​of a physical quantity (e.g., RSSI) received from all neighboring nodes within the last 10 heartbeat cycles). Taking RSSI as an example, let the set of all raw RSSI values ​​(negative values, such as -45, -62, -78...) stored within the current sliding window be denoted as […]. The absolute value of all values ​​in the set is obtained. Find the maximum value in. and minimum value Then, the absolute value of RSSI of the current node's heartbeat is normalized using the Min-Max normalization method. Other physical quantities are normalized in the same way based on their respective sliding windows. Substitute the normalized parameters into the following formula: Among them, CPU_Free is the CPU idle rate, Memory_Free is the remaining memory, Battery_Level is the battery percentage, and RSSI is the signal strength; α, β, γ, and δ are weighting coefficients used to balance the importance of each indicator, satisfying α + β + γ + δ = 1. Specific values ​​can be configured according to the application scenario preferences in the industrial field. For example, in a battery-powered wireless sensing scenario, γ = 0.5 (battery priority); in a computing power-sensitive scenario, α = 0.4 (computing power priority). The system provides a default set of balancing coefficients: α = 0.3, β = 0.2, γ = 0.3, δ = 0.2. The node with the highest score becomes the host node; other nodes automatically join the network as members. This election mechanism ensures that the network coordinator is always the node with the best current resources. 3) When a member node's abnormal MSAD confidence score, output by the MSAD algorithm model, falls into the "uncertainty interval," it initiates a "cooperative verification request" to the host node. Upon receiving this request, the host node, based on the request type and network resource status, instructs one or more other nodes that are complementary in physical location or data type (the "complementarity" criteria include: spatial complementarity: the physical straight-line distance between the cooperating node and the verified node is less than a preset threshold (default 5 meters); modal complementarity: the sensor type of the cooperating node is different from that of the verified node, for example, a vibration node (node ​​A) requests a temperature node (node ​​B) and a current node (node ​​C) to participate in the verification). The host node selects nodes according to a pre-configured cooperative strategy library. The cooperative verification mechanism selects two cooperative nodes by default to balance network overhead while ensuring decision reliability. Users can set the upper limit of the number of participating cooperative nodes in the cooperative strategy configuration interface of the management platform (default no more than 3). At this point, these designated nodes become "cooperative nodes," synchronously collecting data and performing local analysis or distributed inference; each cooperative node feeds back the analysis results to the host node, which then performs information fusion and makes the final decision.

[0019] The present invention discloses a prediction method for a predictive maintenance system for equipment based on multimodal data fusion and edge collaborative decision-making. The specific prediction method is as follows: S1): The intelligent data acquisition card collects multimodal physical signals from industrial equipment; S2): The analysis and operation unit in the core processing unit uses the MSAD algorithm model to perform real-time analysis on the multimodal physical signals collected in S1). The decision fusion layer fuses the data to obtain a quantified anomaly confidence score and fault type label. S3): Based on the set multi-level threshold rules, if the abnormal confidence score is in the uncertainty range between the high and low thresholds, a collaborative verification mechanism request is initiated to the host node through the LoRa networking unit; S4): The host node coordinates relevant collaborating nodes to perform data synchronization and distributed inference; S5): The host node summarizes the local results of each cooperating node, performs data fusion and final decision-making, triggers corresponding control commands, and reports the decision summary to the management platform.

[0020] As can be seen from the above technical solution, the present invention has the following beneficial effects: 1. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making described in this invention integrates a novel AI processing kernel and a dynamic collaborative networking mechanism into an intelligent data acquisition system. It achieves real-time and accurate perception of the status of industrial equipment at the edge through a multimodal perception anomaly detection (MSAD) algorithm designed specifically for edge collaboration and a confidence-based adaptive decision-making mechanism; it realizes self-organization and collaborative computing among multiple nodes through a dynamic host election and task scheduling protocol based on resource perception and reinforcement learning; and finally, it generates accurate control commands and maintenance suggestions through a multi-level decision-making framework for predictive maintenance.

[0021] 2. This invention also utilizes a dual-branch neural network (CNN+GRU) specifically designed for edge collaboration scenarios to achieve deep fusion and uncertainty quantification of multimodal data; it achieves on-demand allocation and joint diagnosis of computing power among multiple nodes through a resource-aware dynamic networking protocol and a trigger-based collaborative verification mechanism; and it generates hierarchical response strategies through configurable multi-level threshold rules. This invention solves the problems of insufficient reliability of single-point decision-making, low utilization of static networking resources, and simplistic decision-making mechanisms in traditional edge solutions. While maintaining the low latency advantage of the edge, it significantly improves the accuracy of fault warning and system reliability through intelligent collaboration among nodes. This invention addresses the lack of intelligent collaboration and ambiguous decision-making mechanisms in current data acquisition systems, significantly improving the accuracy and reliability of fault warning for industrial equipment. Attached Figure Description

[0022] Figure 1 This is a diagram of the overall system architecture in this invention; Figure 2 This is a block diagram of the hardware structure of the intelligent data acquisition card in this invention; Figure 3 This is a flowchart of the data processing and inference process of the MSAD algorithm model in this invention; Figure 4 This is a flowchart of the dynamic self-organizing network and host election process in this invention; Figure 5 This is a flowchart of the single-node edge AI autonomous control process in this invention; Figure 6 This is a schematic diagram of the multi-node collaborative decision-making process in this invention. Detailed Implementation

[0023] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] like Figure 1 The device predictive maintenance system based on multimodal data fusion and edge collaborative decision-making, as shown, includes: an edge side, comprising at least two smart data acquisition cards, which are deployed on the device for data acquisition and control, and are interconnected through an internal LoRa networking unit to form a dynamic Mesh network. One smart data acquisition card is elected as the host node, and the others are collaborative nodes; the host smart data acquisition card communicates with the cloud through a network unit. A unified management platform, comprising a cloud server and user terminals, is provided. The cloud server interacts with host nodes, and the user terminals access the cloud server via the internet for monitoring and management. It also provides AI model management functionality, supporting the remote distribution of trained MSAD algorithm model files as differential upgrade packages to designated intelligent data acquisition cards or device groups. The specific differential upgrade process is as follows: 1) Differential Packet Generation: The cloud server compares the new version of the MSAD algorithm model file with the old version of the model file byte by byte, extracts the difference (Delta) to generate a differential upgrade packet, and calculates its MD5 value; MD5 is a widely used cryptographic hash function that can produce a 128-bit (16-byte) hash value to ensure the integrity and consistency of information transmission. In this embodiment, it is used for verification. 2) Remote distribution: The cloud server splits the differential packet into data blocks and distributes them one by one to the designated target intelligent data acquisition card through the host node; In this embodiment, a mesh network is used, and each device in the network has its own number (operated by the wireless module). Users can also match the number with the corresponding placement location themselves, and upgrade according to the number when upgrade is needed; 3) Verification and Deployment: After the target acquisition card receives the data, it calculates the MD5 value of the received data and compares it with the original MD5 value sent from the cloud to ensure data integrity. After the verification is successful, the acquisition card uses the built-in Bootloader program to merge the differential packet with the local old model, generate a new model, and load and run it. Specifically, it compares a known value with a calculated value to see if they are the same. If they are different, it means there is a problem with the data transmission process, and the data is retransmitted. The specific model merges a bunch of binary data. The update is to send the latest data packet to the lower-level machine. Using differential packets means that the differences between the two data packets are packaged and sent. The lower-level machine receives and verifies that there are no problems before replacing (only the differences are replaced). 4) Safe rollback: If the new model fails to load or causes system instability, the acquisition card will automatically switch back to the old version and report the failure log to the management platform.

[0026] It should be noted that: 1) MD5 is a widely used cryptographic hash function that produces a 128-bit (16-byte) hash value to ensure the integrity and consistency of transmitted information. In this paper, it is used for verification. This can be achieved using existing technology, and will not be elaborated upon here.

[0027] The intelligent data acquisition card performs real-time analysis on locally acquired multimodal sensor data based on the built-in MSAD algorithm model and outputs a quantified anomaly confidence score. When the anomaly confidence score is within the uncertainty range between the first high threshold and the second low threshold preset by the system, a collaborative verification mechanism is triggered. The collaborative verification mechanism is jointly completed by the requesting node, the dynamically elected host node, and one or more collaborative nodes, which synchronously acquire data, perform distributed reasoning and result fusion, and finally form a joint decision at the edge.

[0028] The intelligent data acquisition card described in this embodiment includes a data acquisition unit, a control unit, a core processing unit, a communication unit, and an auxiliary unit. The data acquisition unit, control unit, communication unit, and auxiliary unit are all bidirectionally connected to the core processing unit. The data acquisition unit includes an analog acquisition unit and a digital acquisition unit, and the control unit includes an analog control unit and a digital control unit; The core processing unit is based on the analysis and calculation unit, and the MSAD algorithm model is located within the analysis and calculation unit. The communication unit includes a networking unit and a network unit; the collaboration between the analysis and computing unit and the networking unit jointly supports the MSAD algorithm model and the dynamic collaboration mechanism. The auxiliary unit includes a power supply unit and an indicator unit.

[0029] In this embodiment, preferably, the analysis and computing unit adopts a heterogeneous computing architecture, which includes a multi-core CPU and a dedicated AI acceleration chip for running the Multimodal Perception Anomaly Detection (MSAD) algorithm model.

[0030] It should be noted that the AI ​​acceleration chip is not limited to a specific standalone AI acceleration chip, but can also be used to deploy AI models in NPU, GPU, FPGA or to use a computing unit with integrated AI acceleration units.

[0031] The power supply unit is not limited to an external DC power supply; it can also be powered by a solar panel, a built-in rechargeable battery, or via Power over Ethernet (PoE).

[0032] In this embodiment, the networking unit is used to realize inter-node networking communication and uplink data transmission to the master node. The communication methods used for inter-node networking communication include, but are not limited to, LoRa wireless communication modules, and can also use Wi-Fi Mesh, Zigbee, Bluetooth Mesh or other communication technologies that support self-organizing networks; the uplink data transmission methods to the master node include, but are not limited to, wired Ethernet, wireless Wi-Fi, and cellular networks.

[0033] In this embodiment, the preferred MSAD algorithm model adopts a dual-branch deep neural network structure, including a numerical anomaly perception branch for capturing local feature patterns and a sequence anomaly perception branch for capturing long-term sequence dependencies. The numerical anomaly perception branch adopts a one-dimensional convolutional neural network, and the sequence anomaly perception branch adopts a gated recurrent unit. The MSAD algorithm model performs fusion analysis on the collected multimodal time series data, outputs anomaly confidence scores and fault type labels, and triggers different levels of response according to preset multi-level threshold rules.

[0034] In this embodiment, the device predictive maintenance system based on multimodal data fusion and edge collaborative decision-making includes a multi-level threshold rule that specifically defines: If the abnormal confidence score is greater than the first high threshold, local control will be triggered immediately and a high-risk alarm will be reported. If the abnormal confidence score falls within the "uncertainty interval" between the first high threshold and the second low threshold, the collaborative verification mechanism is activated to request verification from neighboring nodes. If the anomaly confidence score is lower than the second lowest threshold, only the data is recorded and no active response is triggered. Multi-level threshold calibration methods need to be calibrated in conjunction with historical data and are not fixed.

[0035] The system provides a threshold self-learning function: 1): Collect historical data under normal equipment conditions and input it into the MSAD model to obtain the confidence distribution. Set the 95th percentile as the low threshold T_low. 2): Collect known fault injection or historical fault data to obtain the confidence distribution, and set the 5th percentile as the high threshold T_high. The examples given in this specification represent calibration results under specific test conditions only. For different equipment (such as fans, machine tools, pumps), the above calibration process needs to be re-executed in actual applications to adapt to different operating conditions.

[0036] In a preferred embodiment of this invention, the threshold self-learning function adopts an offline (default) + online adaptive approach. During the initial deployment of the equipment or after a change in operating conditions, it calculates the threshold using historical normal data and historical fault data in a single step. and . ( : 95th percentile of confidence level for normal data; (5th percentile of fault data confidence level). This method is suitable for scenarios with relatively stable operating conditions. The update cycle can be monthly or quarterly, and can be manually triggered by maintenance personnel on the management platform. The system can further support online sliding window updates to collect the latest QUOTE data. Dynamically recalculate QUOTE based on daily (e.g., 7 days) normal operating data. Collect recently confirmed fault data and dynamically recalculate. The recommended update frequency is daily or weekly to avoid instability in decision-making due to frequent updates. Online updates should be combined with a manual verification mechanism to prevent abnormal data from contaminating the threshold.

[0037] Threshold processing is as follows: Silent recording (≤ 0.7, i.e., below) During this phase, data is simply stored in a local cache or edge log without reporting or intervention, and is used for subsequent model iterations; collaborative validation (0.7 ~ 0.9) to When a high-risk alarm occurs, the network unit requests nearby nodes to synchronously collect and jointly infer data, and the host node then merges the data for a secondary decision; high-risk alarm + local control (≥ 0.9, i.e., higher than...). When this occurs, the local actuator (such as relay cut-off or buzzer alarm) is immediately triggered, and the cloud is simultaneously notified and an alarm is pushed.

[0038] like Figure 3 The data processing flow of the MSAD algorithm model shown in this embodiment is as follows: 1) Data preprocessing and alignment: After receiving the raw data collected by the data acquisition unit, a joint denoising algorithm based on wavelet transform is used to filter the high-noise industrial data: The 'db4' wavelet basis function is used to perform 5-level discrete wavelet decomposition on the raw signal, and the high-frequency detail coefficients are denoised using a soft threshold function. After denoising, the dynamic time warping (DTW) algorithm is used to align the time-series asynchronous data between multimodal sensor data. During alignment, the bending window size is set to 10, and the distance measurement method is Euclidean distance to ensure the synchronization of different modal data (such as vibration and current) on the time axis. The 5-level wavelet decomposition includes both detail coefficients and approximation coefficients, as detailed below: The first to fifth layers output high-frequency detail coefficients (d1~d5) and the fifth layer outputs low-frequency approximation coefficients (a5). Noise reduction typically involves soft thresholding of high-frequency detail coefficients, while preserving or slightly reducing approximation coefficients (which represent the main trend of the signal and are usually not subject to strong noise reduction). After noise reduction, wavelet reconstruction must be performed, that is, the processed detail coefficients and approximation coefficients are restored to the time domain signal through inverse wavelet transform, which is used as the input for subsequent DTW alignment; The soft thresholding process uses the classic wavelet denoising method proposed by Donoho, which is a mature technology in the field of signal processing and will not be described in detail here. Original wavelet detail coefficients For threshold values, a common threshold value is typically used. , The standard deviation of the noise. The original signal length; These are the processed wavelet coefficients; The alignment process is as follows: DTW is used to align time-series data (such as vibration and current) acquired by two different sensors, eliminating timeline misalignment caused by different sampling start points, transmission delays, or physical response delays. Alignment criteria: To minimize the sum of distances between corresponding points when two sequences are "distorted" on the time axis.

[0039] Alignment process (core steps): Construct a distance matrix for two sequences Calculate each point pair The distance (Euclidean distance) forms matrix; Dynamic Programming Cumulative Distance Definition Cumulative Distance ,from Begin calculating cell by cell; Finding the optimal path (backtracking) from Start by selecting the cell with the smallest accumulated distance from the previous step in reverse order, until you return to the starting position. This yields the optimal alignment path; The alignment output aligns corresponding points in two sequences according to the path, which is used for subsequent feature extraction. Using Euclidean geometry allows for a unified scale: multimodal data (such as vibration amplitude and current amplitude) have different physical dimensions, but in DTW they are only used as a "measure of difference" and do not depend on physical units; The calculation is simple: Euclidean distance is an L2 norm, which is efficient for calculation in dynamic programming; Preserving the original form: Compared with Manhattan distance or cosine distance, Euclidean distance is sensitive to amplitude differences and is suitable for capturing abnormal fluctuations; 2): Multi-scale feature extraction, which utilizes the parallel computing capabilities of the AI ​​acceleration chip in the analysis and computing unit to simultaneously extract multi-scale feature values ​​of time domain features, frequency domain features, and time-frequency domain features from the preprocessed data. The system then fuses the extracted feature values ​​into a single feature vector. Specifically, this includes: time-domain feature extraction (such as calculating the root mean square, kurtosis, peak factor, etc. of vibration signals), frequency-domain feature extraction (i.e., calculating the spectrum of the original signal through FFT and extracting characteristic frequencies, such as the vibration frequency amplitude under bearing operation), and time-frequency-domain feature extraction (i.e., performing wavelet packet transform (WPT) and calculating the energy entropy of each frequency band). To eliminate dimensional differences between different features due to their different physical meanings (e.g., kurtosis is dimensionless, while energy entropy has units), the Min-Max normalization method can be used to process the data; the aim is to uniformly scale all features to the numerical range [0,1], as shown in the following formula: in, These are the original eigenvalues; These are the minimum and maximum values ​​of the original feature values ​​in the training set, respectively, used to ensure the consistency of the transformation; These are the normalized eigenvalues ​​(dimensionless quantities). The above steps yield the time-domain features, frequency-domain features, and normalized feature vectors of the time-frequency features, respectively. Considering the different contributions of time-domain features, frequency-domain features, and time-frequency features to fault diagnosis, a weighted fusion strategy is adopted. Fusion weights are set (weights can be learned through training or domain knowledge, and must satisfy normalization constraints, as shown in the following formula): The core fusion formula first weights the features from each domain, then performs vector concatenation to preserve all feature dimensions and avoid information loss. in, These represent the weighting coefficients for time-domain, frequency-domain, and time-frequency-domain features, respectively; ⊕: indicates vector concatenation operation; weighting coefficients for time-domain, frequency-domain, and time-frequency-domain features. The determination involves two stages: initialization and dynamic adjustment. Initialization stage: Based on historical fault data, principal component analysis (PCA) is used to calculate the contribution rate of each domain feature to fault classification. Assume the variance contribution rates of the first three principal components extracted by PCA are as follows: Then the weights of each domain can be initialized as follows: For example, if calculated The weights are then initialized as follows: .

[0040] Dynamic Adjustment Phase: During system operation, the weights are fine-tuned online using gradient descent. The loss function is defined as the cross-entropy between the model's predicted confidence and the actual fault label. The gradient of the loss function with respect to the weights of each domain is calculated through backpropagation, using the learning rate... The weights are updated gradually to make the fused feature vector more closely resemble the fault feature distribution under the current operating conditions.

[0041] The following specific calculation example illustrates this: Assume the normalized feature vectors and weights are as follows: Temporal characteristics: (3-dimensional), weights ; Frequency domain characteristics: (2-dimensional), weights ; Frequency domain characteristics: (1-dimensional), weights ; The calculation process is as follows: 1. Weighted processing: 2. Vector concatenation: fused feature vector 3) Anomaly detection and inference: The fused single feature vector is input into the MSAD algorithm model for inference. Specifically, a one-dimensional convolutional neural network of the numerical anomaly perception branch is used to capture local feature patterns, and a gated recurrent unit of the sequence anomaly perception branch is used to capture long-term sequence dependencies. The outputs of the numerical anomaly perception branch and the sequence anomaly perception branch are summarized in the decision fusion layer to form a fused comprehensive feature. The comprehensive feature is fed into the fully connected layer to finally generate a comprehensive anomaly confidence score (0-1) and a preliminary fault type label. This anomaly confidence score is the direct basis for triggering different levels of response and collaborative verification. The specific network structure of the MSAD algorithm model is as follows: Numerical Anomaly Awareness Branch (CNN): It contains 3 one-dimensional convolutional layers. The first layer has 64 kernels, size 3, and stride 1; the second layer has 128 kernels, size 3, and stride 1; the third layer has 256 kernels, size 3, and stride 1. Each convolutional layer is followed by a ReLU activation function and a max pooling layer (pooling size 2) to extract local feature patterns; Sequence Anomaly Detection Branch (GRU): Contains two stacked GRU layers, each with 128 hidden neurons, used to capture long-term sequence dependencies; Decision Fusion Layer: Concatenates the output of the CNN branch (flattened) with the output of the GRU branch at the last time step to obtain a comprehensive feature vector (dimension 1024); Output Layer: The fused comprehensive features are fed into a fully connected layer, passing through two fully connected network layers (the first layer has 512 neurons, ReLU activation; the second layer uses the Sigmoid activation function) to finally generate a comprehensive anomaly confidence score (0-1), and simultaneously a softmax layer is connected in parallel to output a preliminary fault type label; Model Training Parameters: The model is trained using the Adam optimizer, with an initial learning rate of 0.001, a loss function that is the weighted sum of binary cross-entropy (for confidence output) and classification cross-entropy (for fault classification), a training batch size of 32, and 100 epochs.

[0042] The MSAD algorithm model has the following requirements for the training dataset: 1. Data source: Data collected by sensors of the same model as the target device, including multimodal signals (vibration, temperature, current, etc.); 2. Data composition: Normal samples (70%~80%), known fault samples (20%~30%), and the fault types should cover common fault modes; 3. Labeling requirements: Each sample segment needs to be labeled: 0 (normal), 1~K (specific fault type), and an anomaly confidence soft label (optional); 4. Data length: Each sample is a continuous time segment with a fixed length, used to capture local features and long-term dependencies; 5. Sample size: The training set should have no less than 10,000 samples, and the validation set should have no less than 2,000 samples (adjustable depending on device complexity). Timing window length (Refers to the continuous time length covered by each model inference), defaulting to 10-30 seconds (for rotating machinery) or 2-5 times the fault feature cycle. The specific value depends on the equipment's dynamic response speed and sampling frequency (e.g., a vibration signal sampling rate of 2kHz, a 20-second window corresponding to 40,000 sampling points, compressed into a fixed-length sequence after downsampling and feature extraction); feature vector input length. (Refers to the time steps of the input model). After preprocessing, the original waveform is mapped to a feature vector (such as mean, peak value, spectral energy, etc.) through feature extraction. Multiple consecutive windows form a length of... The temporal feature sequence. T is typically set to 50-200, corresponding to historical information of several seconds to tens of seconds, ensuring that GRU can capture long-term dependencies while controlling model complexity. Feature Dimension The feature dimension of each time step is determined by the fusion of multimodal features, and is usually between 8 and 128.

[0043] In the preferred implementation of this example, the CNN is used to capture and process local numerical anomalies in the sensor (i.e., single points, such as instantaneous peaks, sudden changes, etc.); the GRU is used to capture sequence dependencies (such as data drift, periodic fluctuations, or anomalies) during long-term operation of the device; the core function of the decision fusion layer is to weightedly fuse the local features and sequence features output from the two branches, ultimately generating a comprehensive anomaly confidence score and fault label. The output of the numerical anomaly perception branch (CNN) is shown below. and the output of the sequence anomaly sensing branch (GRU) Instead of using fixed weighting coefficients during fusion, an attention mechanism is introduced to dynamically generate fusion weights; the specific derivation is as follows: a): will and Each of these is mapped to a score scalar through a fully connected layer: in For a trainable weight matrix, This is a bias term.

[0044] b): Normalize the scores using the Softmax function to obtain the final fusion weights a and b: Therefore, a + b = 1 is automatically satisfied.

[0045] c): Final integrated characteristics in, This represents the integrated characteristics after fusion; This represents the local features of the numerical anomaly perception branch output; This represents the temporal characteristics of the output of the sequence anomaly detection branch; It should be noted that the weight coefficients a and b are obtained through model training and need to be adjusted according to the fault characteristics of different industrial equipment to reflect the difference in contribution between local anomalies and sequence anomalies, and to meet the requirements. ; In this way, the model can automatically assign higher weights to more important branches based on the characteristics of the current input data during inference, thus achieving adaptive feature fusion.

[0046] 4): Adaptive collaborative decision-making mechanism: The system performs different operations through adaptive decision-making based on preset multi-level threshold rules.

[0047] The core of the multi-level threshold rule lies in defining an "uncertainty interval," meaning the system does not trigger alarms for all anomalies but makes decisions according to preset multi-level threshold rules. For example, if the anomaly confidence level is ≥0.9, local control (such as shutdown) is immediately triggered and a high-risk alarm is reported. If the anomaly confidence level is between 0.7 and 0.9, a collaborative verification mechanism is initiated, which requests neighboring nodes to synchronously collect relevant data and perform local verification through the network unit. If the anomaly confidence level is ≤0.7, the system only logs the data and does not trigger an active response, continuing to observe. This mechanism ensures that system resources are efficiently used for scenarios that most require joint judgment. Preferred implementation in this embodiment: control commands can use JSON or binary protocols; reset can be manual (manually reset by maintenance personnel after confirming safety through the management platform or physical button; automatic reset is prohibited); automatic reset (only applicable to low-risk scenarios) can be performed under the following conditions for scenarios that only trigger collaborative verification and are ultimately determined to be "false alarms": The collaborative verification result was determined to be "no fault". Subsequent Continuous The anomaly confidence level for each (e.g., 10) time window is lower than [a certain value]. ; The equipment operating parameters have returned to the normal range.

[0048] The reset mechanism prioritizes safety, requiring manual confirmation for high-risk actions; non-critical states can be automatically restored to avoid excessive operational burden.

[0049] In this embodiment, the intelligent data acquisition card uses a resource-aware distributed host election algorithm to elect a host node. During the host node election process, the resource status of each node is exchanged through heartbeat packets, and the host node is elected based on a dynamic scoring function. The host node then coordinates the member nodes to perform data sharing and distributed collaborative computing, as detailed below: 1) Initial Network Setup and Status Broadcast: After powering on, each node enters network maintenance mode and broadcasts a "heartbeat" packet containing its own device ID, computing power (CPU utilization, memory remaining), power status, and signal strength (RSSI) every T seconds (T is set to 30 seconds by default, and the period can be configured remotely through the management platform). 2) Dynamic Host Node Election: After a node receives a heartbeat packet from another node, it will not immediately trigger a re-election to avoid network jitter. Each node maintains an election timer (the default duration is 3 times the heartbeat cycle, i.e., 90 seconds). When the timer expires, the node collects the latest heartbeat information of all neighboring nodes received within that window period, substitutes it into the scoring function below to calculate the score, and the node with the highest score is identified as the host node. If no host node is specified, all nodes follow this rule, and the node with the highest score becomes the host node. Other nodes automatically join the network as members. Since CPU_Free (percentage), Memory_Free (MB), Battery_Level (percentage), and RSSI (dBm, usually negative, absolute value must be used for calculation) have different units, they need to be Min-Max normalized before being substituted into the scoring function to uniformly map them to the [0, 1] interval. Specific formula: in, Let the minimum and maximum values ​​of this physical quantity be defined within the network during the most recent time window (where each node maintains a sliding window of fixed length L (e.g., L=10) to store the raw values ​​of a physical quantity (e.g., RSSI) received from all neighboring nodes within the last 10 heartbeat cycles). Taking RSSI as an example, let the set of all raw RSSI values ​​(negative values, such as -45, -62, -78...) stored within the current sliding window be denoted as […]. The absolute value of all values ​​in the set is obtained. Find the maximum value in. and minimum value Then, the absolute value of RSSI of the current node's heartbeat is normalized using the Min-Max normalization method. Other physical quantities are normalized in the same way based on their respective sliding windows. Substitute the normalized parameters into the following formula: Here, CPU_Free represents the CPU idle rate, Memory_Free represents the remaining memory, Battery_Level represents the battery percentage, and RSSI represents the signal strength. α, β, γ, and δ are weighting coefficients used to balance the importance of each indicator, satisfying α + β + γ + δ = 1. The system provides a default set of balancing coefficients: α = 0.3, β = 0.2, γ = 0.3, and δ = 0.2. The node with the highest score becomes the host node; other nodes automatically join the network as members. This election mechanism ensures that the network coordinator is always the node with the best current resources. It should be noted that the specific values ​​of α, β, γ, and δ can be configured according to the application scenario preferences in the industrial field. For example, in a battery-powered wireless sensing scenario, γ can be set to 0.5 (battery priority); in a computing power-sensitive scenario, α can be set to 0.4 (computing power priority). The rules for determining the values ​​for application scenarios are as follows: 1. Balanced Mode (Default Scenario): Suitable for general industrial environments with balanced mains power supply, computing power, and communication resources (Values: 0.3, 0.2, 0.3, 0.2); 2. Computing Power Priority Mode: Used for scenarios requiring high-intensity edge inference (such as real-time analysis of high-frequency data). The device uses mains power, and battery power is ignored, emphasizing CPU and memory (Values: 0.5, 0.3, 0.0, 0.2); 3. Energy Saving Priority Mode: Used for battery-powered wireless sensor nodes. Maximizes node endurance with high power weight and reduces communication overhead weight (Values: 0.2, 0.1, 0.6, 0.1); 4. Link Sensitive Mode: Used for factory environments with unstable wireless network signals and susceptibility to interference. Prioritizes communication quality, increases signal strength weight, and ensures stable communication for the master node (Values: 0.2, 0.2, 0.2, 0.4); 5. Lightweight Deployment Mode: Used for resource-constrained simple monitoring nodes (such as those only collecting temperature). Low computing power requirements, emphasizing power and CPU, and de-emphasizing memory. (Values: 0.2, 0.2, 0.2, 0.4); 6. Custom mode: Users can adjust the settings according to their actual needs; 3) When a member node's abnormal MSAD confidence score, output by the MSAD algorithm model, falls into the "uncertainty interval," it initiates a "cooperative verification request" to the host node. Upon receiving this request, the host node, based on the request type and network resource status, instructs one or more other nodes that are complementary in physical location or data type (the "complementarity" criteria include: spatial complementarity: the physical straight-line distance between the cooperating node and the verified node is less than a preset threshold (default 5 meters); modal complementarity: the sensor type of the cooperating node is different from that of the verified node, for example, a vibration node (node ​​A) requests a temperature node (node ​​B) and a current node (node ​​C) to participate in the verification). The host node selects nodes according to a pre-configured cooperative strategy library. The cooperative verification mechanism selects two cooperative nodes by default to balance network overhead while ensuring decision reliability. Users can set the upper limit of the number of participating cooperative nodes in the cooperative strategy configuration interface of the management platform (default no more than 3). Only then do these designated nodes become "cooperative nodes," synchronously collecting data and performing local analysis or distributed inference. Each cooperative node feeds back the analysis results to the host node, which then performs information fusion and makes the final decision. The specific mode of the distributed inference is as follows: The host node splits the computationally intensive fully connected layer task in the MSAD algorithm model according to the idle computing power ratio of each cooperating node. The cooperating nodes only need to compute the subset of neurons assigned to them and return the intermediate results to the host node for aggregation, thereby achieving parallel acceleration. The spatial computing power ratio is split as follows: 1. First, the idle computing power ratio is obtained. Each node periodically reports its resource status through heartbeat packets, including: CPU idle rate (calculated in real time through system load), remaining memory, and current task queue length (reflecting whether it is busy). The host node maintains a node resource table and obtains the latest status before each distributed inference.

[0050] 2. Formula for calculating computing power ratio Assume the total number of neurons in the fully connected layer is The set of collaborative nodes participating in the computation is ,node The weight of idle computing power is Then the node The number of neurons assigned is: in It can be obtained by weighting multiple indicators, for example: This is a weighting coefficient, which can be configured according to the scenario (e.g., it can be increased in scenarios sensitive to computing power). Each indicator needs to be normalized using Min-Max to eliminate the influence of dimensions.

[0051] 3. Parameters are set to limited computing power (e.g., ... The calculation is performed using the smallest allocation unit to prevent fragmentation (e.g., setting a minimum number of neurons to be allocated (e.g., 8, 16, or 32), with any insufficient part being handled by the host node or merged into other nodes), and a dynamic adjustment cycle is adopted (recalculated before each collaborative inference, without long-term caching). 4. Synchronization Mechanism Task distribution: The host node distributes the split computational tasks (including subsets of neuron weights and input vectors) to each collaborating node through the networking unit; Computation execution: Each collaborative node independently completes the forward computation and generates an intermediate result vector (i.e., a part of the output of this layer). Result retrieval: The collaborating node returns the intermediate results to the host node, which then reassembles them in the original order to complete the full calculation of this layer and continues to the subsequent network layers. 5. Timeout and Fault Tolerance Handling Timeout threshold: The default timeout is set to 2 to 3 times the normal inference time of the node (for example, if the local inference takes 100ms, the timeout is set to 300ms). Timeout handling: If a node fails to return a result within the specified time, the host node will dynamically adjust the task allocation, reassigning the unfinished part of the task to other online nodes or calculating it itself. If timeout nodes occur consecutively, they are marked as "computing power unavailable". Subsequent distributed inference will temporarily exclude the node until its heartbeat recovers and its resource status returns to normal. Partial failure tolerance: Distributed inference can be completed as long as the number of participating nodes is ≥ 2; if only a single node remains, it will fall back to local inference to ensure system availability.

[0052] In this embodiment, the reliability of its decision-making is ensured through four levels: mechanism design, data complementarity, decision fusion, and configurable strategies. Specifically: Mechanism level: trigger-based collaboration to avoid invalid verification; Uncertainty interval triggering: Collaboration is only initiated when the MSAD abnormal confidence level of a single node falls between the high and low thresholds, thus avoiding the abuse of network and computing power and concentrating collaborative resources on the "fuzzy state" that truly requires confirmation from multiple sources. Data level: Spatial complementarity + Modal complementarity Spatial complementarity: The physical straight-line distance between the cooperating node and the verified node is less than a preset threshold (5 meters by default), ensuring that they perceive the associated status of the same device or the same physical area, and avoiding the introduction of irrelevant noise.

[0053] Modal complementarity: Selecting different types of sensors (such as vibration, temperature, and current) to participate in the verification utilizes the redundancy and orthogonality between multiple physical quantities. For example, if vibration anomalies are accompanied by temperature anomalies or current harmonic distortion, the reliability of the fault is greatly improved; if there is only a single modal anomaly while other modes are normal, it may be due to local interference or occasional sensor drift. Decision-making level: Host nodes perform multi-source information fusion The host node does not simply collect anomaly judgments from each cooperating node, but rather fuses the local features of multiple nodes (such as temperature rise rate, current spectrum features, and vibration time-frequency domain features) and uses weighted voting or lightweight fusion models (such as logistic regression and DS evidence theory) to output the final decision.

[0054] This fusion approach can effectively suppress false alarms from single nodes and improve the accuracy and robustness of joint diagnosis.

[0055] At the strategy level: User-configurable to adapt to different reliability requirements. Users can set the upper limit of the number of nodes participating in collaboration on the management platform (the default is no more than 3), and can configure the selection strategy for collaborative nodes (such as giving priority to nodes with sufficient power, idle computing power, and good signal quality).

[0056] For critical equipment, it is mandatory to require at least two different modalities of nodes to participate in the verification process to further improve reliability.

[0057] It's important to note that upon receiving a request, the host node doesn't simply collect data. Instead, based on the request type and the resource status of neighboring nodes, it instructs one or more neighboring nodes to synchronously collect relevant data. It may also split the AI ​​model computation tasks required for verification onto the idle computing power of each node for parallel execution (distributed inference). Finally, the host node aggregates the partial conclusions from each node, performs data fusion, and makes the final decision. This achieves collaboration across the entire "perception-computation-decision" chain.

[0058] Example 2 The device predictive maintenance system based on multimodal data fusion and edge collaborative decision-making described in this embodiment is the same as in Embodiment 1. Preferably, in this embodiment, the inter-node networking method uses a LoRa wireless communication module and runs a resource-aware distributed host election algorithm. Through the networking protocol and the resource-aware distributed host election algorithm, a dynamic, self-healing Mesh network is constructed. The specific implementation of the "self-healing" mechanism is as follows: 1) Fault determination: If a member node does not receive a broadcast message from the host node for three consecutive heartbeat cycles (i.e., 90 seconds), or detects that the communication error rate with the master node is consistently higher than 10%, then the current link or the master node is determined to be faulty. 2) Self-healing action: The node that determines the fault immediately broadcasts a "re-election request." All online nodes clear their original host records and re-execute the dynamic host election process. The newly elected host node rebuilds the network topology and incorporates the faulty node (if online) or the newly joined node into the new network. It should be noted that the model and thresholds exist only in individual devices and are unrelated to the network, because each slave has the potential to become a master, and a master may also become a slave; the network only serves as a data transmission medium.

[0059] The resource-aware distributed host election algorithm communicates node resource status via heartbeat packets and elects host nodes based on a dynamic scoring function. The host nodes then coordinate member nodes for data sharing and distributed collaborative computing.

[0060] The logical flow of the networking unit establishing the network in this embodiment is as follows: First, each node (master node + cooperating node) is powered on and initialized. After initialization, it periodically broadcasts a "heartbeat" packet (the packet contains information such as its own device ID, computing power, battery level, and signal strength). Each node receives and parses the heartbeat packets of other nodes, and calculates its own score and the scores of other nodes in the network according to a preset dynamic scoring function. Each node then determines whether its own score is the highest in the network. If the determination is yes, the node declares itself as the host node; if the determination is no, it joins the network built by the highest-scoring node as a member node; the network enters a stable working state, and the host node begins to coordinate communication and tasks.

[0061] The unified management platform is not limited to the form of APP, web page or desktop software, but can also be integrated into other industrial monitoring systems or platforms or embedded in sub-applications of APP; The policy-based configuration function of the unified management platform includes: a) Differential AI model upgrade: In addition to the system's preset AI models, users can also upload MSAD model files trained on the cloud platform for specific device types. The unified management platform can securely distribute the differential upgrade package to designated devices or device groups, enabling remote deployment and updates of AI models. It should be noted that differential AI model upgrade is an incremental update mode, which means that without completely replacing the entire model, only the changed parts of the model parameters are transmitted and applied, thereby achieving model updates and optimizations; similar to the incremental update package of a mobile phone system, rather than downloading the entire new system. b) Graphical configuration of collaborative strategies: Through a cross-platform graphical interface, users can configure multi-level threshold rules and parameters for collaborative verification mechanisms. These parameters include the confidence interval for triggering collaborative verification and the combination of nodes participating in the collaboration. During use, users can deeply customize the aforementioned multi-level threshold rules and parameters for collaborative verification mechanisms, such as setting the confidence interval for triggering collaborative verification and the combination of nodes participating in the collaboration. The "deep customization" mentioned above is not completely unrestricted free input, but rather configuration within a pre-set rule framework to ensure system stability, security, and interoperability. The specific rules are as follows: First, the hierarchy is divided into: Basic mode (ordinary operation and maintenance personnel, only modifying thresholds (slider) and number of collaborative nodes); Advanced mode (expert users, all parameters (including weight coefficients, fusion algorithms, and distributed inference parameters)); and API mode (system integration developers, who can configure in batches via REST API and support integration with third-party operation and maintenance systems). Second, the parameter value range and type are constrained to avoid inputting abnormal parameters. Next, you need to configure the consistency verification rules, as follows: Threshold monotonicity: System-mandated verification T low <T high Otherwise, the message "The lower threshold must be less than the higher threshold" will be displayed.

[0062] Node Existence: If the number of configured collaborative nodes exceeds the actual number of nodes in the current device group, the system will automatically retrieve the actual number of nodes and provide a prompt.

[0063] Modal Availability: If "Modal Complementarity" is configured but there are not enough nodes of different modes in the device group, the system will prompt "Insufficient available modes, it is recommended to adjust the strategy".

[0064] Finally, configure the rules for activation and rollback as follows: Canary rollout: Critical configurations (such as thresholds and the number of collaborative nodes) can be rolled out in a canary manner by device group to avoid risks caused by simultaneous changes across the entire network; One-click rollback: Each configuration change automatically backs up the previous version, and users can roll back to any historical configuration with one click; Configuration audit: All configuration change records the operator, time, and content of the change for easy traceability; c) Visualized collaborative decision-making: It provides a visual interface for collaborative decision-making, which can not only display real-time data and historical playback data, but also provide a visual interface for root cause analysis of faults, display the results of multi-node data correlation analysis, and the prediction curve of remaining useful life (RUL).

[0065] It should be noted that the graphical configuration of collaborative strategies and the visualization of collaborative decisions are both implemented through the operation module deployed on the cloud server. During the graphical configuration of collaborative strategies, the operation interface has input boxes and selectors, which can be set and selected by the user.

[0066] The unified management platform (Web / Desktop) provides a visual configuration interface. Users can complete the collaboration strategy configuration by following these steps: Step 1): Access the policy configuration module • Log in to the unified management platform, and select “Collaborative Policy Management” → “Policy Configuration” in the left navigation bar; The top of the interface displays the currently selected equipment group (e.g., "Wind Turbine Unit_01"), and supports filtering by region and equipment type; Step 2): Configure multi-level threshold rules Configuration items Operation method illustrate low threshold Slide bar (0~1) + Numeric input box The default value is 0.7, and the recommended range is 0.5~0.85. High threshold Slide bar (0~1) + Numeric input box The default value is 0.9, the recommended range is 0.85~0.98, and it must be greater than 0.98. Threshold calibration method Drop-down selection: Offline calibration / Online sliding window Online mode requires setting the window duration (default 7 days). High-risk alarm actions Checkbox groups: Local shutdown, buzzer alarm, cloud push, SMS notification Multiple selections allowed Silent recording of actions Checkboxes: Local storage, periodic reporting Select all by default Step 3): Configure collaborative verification mechanism parameters Configuration items Operation method illustrate Number of cooperating nodes Number input box (1~5) The default value is 2, and the upper limit can be set (the default value does not exceed 3). Spatial complementary radius Numeric input box (unit: meter) The default distance is 5 meters, used to filter physically neighboring nodes. Modal complementarity strategy Multi-select dropdown Optional modes include vibration, temperature, current, voltage, and acoustic signature; the system automatically selects from different modes. Collaboration timeout Numeric input field (unit: milliseconds) The default timeout is 3000ms; if it exceeds this, the host node will make a fallback decision. Fusion Algorithm Drop-down selection: Weighted Voting / DS Evidence Theory / Lightweight Neural Network Default weighted voting Step 4): Configure collaborative verification mechanism parameters Configuration items Operation method illustrate Enable Distributed Reasoning switch button Enabled by default Minimum number of neurons allocated Number input box The default value is 32, to prevent fragmentation. Idle computing power weight coefficient Three sliders (CPU / Memory / Queue Length) The default values ​​are α=0.5, β=0.3, and γ=0.2. The normalized weights are displayed in real time. Step 5): Preview and Publish Click the "Preview Strategy" button, and the system will display the current configuration in the form of a natural language summary (such as "When the anomaly confidence is between 0.7 and 0.9, select 2 neighboring nodes (including at least 1 different modality) for collaborative verification, with a timeout of 3 seconds").

[0067] After confirming that everything is correct, click "Publish". The configuration command will be broadcast to all member nodes via the cloud server → host node. After receiving the command, the nodes will update their local policy parameters and return confirmation.

[0068] In this embodiment, a preferred implementation method is provided: the RUL prediction function provided by the system is based on the degradation characteristics output by the MSAD model and historical fault data, and uses a similarity matching model or a time-series prediction model to estimate the remaining lifetime. The specific method is as follows (which can be used as a supplement to the preferred embodiment): 1. Similarity-Based Matching (SBM) Applicable scenarios: Sufficient historical fault data (complete lifecycle data of at least 10 similar devices). step: Construct a health baseline library: Normalize the degradation feature sequence of historical equipment from health to failure by time to form a "feature-lifetime" trajectory library.

[0069] Feature alignment: For the latest degradation features of the current device, the DTW algorithm is used to align them with each trajectory in the benchmark library to find the K most similar trajectories.

[0070] Lifetime weighted: The remaining lifetime of these K trajectories is weighted by similarity to obtain the current device's RUL prediction value.

[0071] Confidence Interval Output: Outputs the 50%~90% confidence interval of RUL, facilitating operational and maintenance decision-making. 2. Based on time series prediction models (LSTM / Transformer) Applicable scenarios: When historical fault data is insufficient, but continuous health degradation monitoring data is available. step: Offline training: Train an LSTM network with an encoder-decoder structure using the historical degradation feature sequence (normal → faulty). Input the degradation features of the most recent L time steps and output the feature predictions for the next H time steps.

[0072] Threshold determination: Input the predicted feature vector into the classification head of the MSAD model to determine the time point when the "failure threshold" is reached, and then infer the remaining lifespan.

[0073] Online updates: The forecast results are updated on a rolling basis and a trend curve is output whenever new monitoring data is received.

[0074] Example 3 This embodiment presents a prediction method for a predictive maintenance system based on multimodal data fusion and edge collaborative decision-making. The prediction method is based on the predictive maintenance system based on multimodal data fusion and edge collaborative decision-making described in Embodiment 1, and its specific prediction method is as follows: S1): The intelligent data acquisition card collects multimodal physical signals from industrial equipment; S2): The analysis and operation unit in the core processing unit uses the MSAD algorithm model to perform real-time analysis on the multimodal physical signals collected in S1). The decision fusion layer fuses the data to obtain a quantified anomaly confidence score and fault type label. S3): Based on the set multi-level threshold rules, if the abnormal confidence score is in the uncertainty range between the high and low thresholds, a collaborative verification mechanism request is initiated to the host node through the LoRa networking unit; S4): The host node coordinates relevant collaborating nodes to perform data synchronization and distributed inference; S5): The host node summarizes the local results of each cooperating node, performs data fusion and final decision-making, triggers corresponding control commands, and reports the decision summary to the management platform.

[0075] Example 4 In this embodiment, the equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making makes single-node and multi-node collaborative decision-making processes as follows: like Figure 5The flowchart shown illustrates the independent workflow of a smart data acquisition card 1 without relying on network collaboration, i.e., the single-node edge AI autonomous control process is as follows: First, the device is powered on and initialized. After initialization, each acquisition unit acquires data according to preset parameters. The collected data is fed into the MSAD algorithm model for real-time analysis (corresponding to...) Figure 3 The process involves determining the confidence level of an anomaly. If the confidence level is ≥0.9, local control commands (such as shutdown or alarm) are triggered, and a high-risk alarm is reported to the cloud. If the confidence level is between 0.7 and 0.9, a collaborative verification mechanism is initiated (see below for details). Figure 6 If the confidence level is ≤0.7, only local data is recorded, and the process ends.

[0076] like Figure 6 As shown, this diagram illustrates a multi-node collaborative decision-making workflow. Taking wind turbine bearing fault diagnosis as an example, the diagram uses a sequence diagram to demonstrate... Figure 5 The detailed process of the collaborative verification mechanism is as follows: the simulation object consists of node A (vibration), node B (temperature), node C (current), host node, and cloud platform; When node A detects a medium confidence anomaly, it sends a "cooperative verification request" to the host node; Upon receiving the request, the host node immediately sends an "instruction" to nodes B and C, requesting them to synchronously collect temperature and high-frequency current data. Nodes B and C perform data acquisition and utilize their own computing power for local analysis (e.g., B calculates the temperature rise rate, and C performs current spectrum analysis). It should be noted that the intelligent data acquisition card uses a processor with AI computing power, and the AI ​​model can be updated in real time through a cloud server during subsequent work. Therefore, the intelligent data acquisition card can use its own computing power to perform calculations on the collected data. Specifically: 1. Default temperature calculation cycle: 10 seconds (can also be adjusted through the "Collaborative Strategy Configuration" interface of the management platform); temperature sampling rate (1 point per second, industrial temperature changes slowly, high-frequency sampling is meaningless), analysis window (60 seconds, covering a sufficiently long temperature change trend, excluding instantaneous fluctuations), sliding step size (10 seconds, (consistent with the calculation cycle), taking the data from the most recent 60 seconds for each calculation); linear regression method is used to calculate the temperature rise rate; Anomaly detection: Compare the calculated temperature rise rate with a preset threshold (e.g., 0.5℃ / minute). If the threshold is exceeded, a "temperature anomaly" flag and the specific rate value are output for the host node to make a fusion decision. 2. Default current calculation period: 30 seconds (current signals change relatively quickly, but the calculation of spectrum analysis is large, so it should not be too frequent. It can also be adjusted through the "Collaborative Strategy Configuration" interface of the management platform); Current sampling rate (2 kHz, covering the fault characteristic frequencies of industrial motors (such as 2nd harmonics and sidebands). Analysis window (0.5 seconds, ensuring frequency resolution = sampling rate / number of window points, sufficient to distinguish common fault characteristics) FFT points (1024 (padding to powers of 2 to improve spectral smoothness), sliding step size (30 seconds (consistent with the analysis cycle), re-acquiring the latest 0.5 seconds of data for each calculation, without overlap; using Fast Fourier Transform (FFT) + feature frequency extraction); The specific abnormality determination is as follows: If the second harmonic amplitude exceeds 5% of the fundamental frequency amplitude, it is determined as "rotor imbalance tendency"; If the sideband amplitude exceeds 2% of the fundamental frequency amplitude, it is judged as "rotor bar breakage tendency"; If the high-frequency energy exceeds the normal baseline by 20%, it is judged as "early bearing damage tendency".

[0077] Nodes B and C will feed the analysis results back to the host node; The host node makes a final decision by combining all information (the original request from node A + the verification results from nodes B / C) and confirms that the risk of failure is extremely high. The host node sends a "control command" to node A (or the actuator), such as suggesting a shutdown. Simultaneously, the host node uploads a complete diagnostic report to the cloud platform, which then pushes a high-level alarm message to the administrator. It should be noted that all nodes are in a mesh network, and communication must be transmitted over the network. The command priority rules can use the default settings or be customized according to the administrator's needs. The reset process has already been described in terms of manual and automatic resets and will not be repeated here. Example 5 This embodiment describes a prediction method for a predictive maintenance system based on multimodal data fusion and edge collaborative decision-making. This prediction method is based on the predictive maintenance system based on multimodal data fusion and edge collaborative decision-making described in Embodiment 1. The specific prediction method is as follows: S1: Intelligent data acquisition card 1 acquires multimodal physical signals from industrial equipment; S2: The analysis and calculation unit in the core processing unit uses the MSAD algorithm model to perform real-time analysis on the multimodal physical signals acquired in S1. The decision fusion layer fuses the data to obtain quantified anomaly confidence scores and fault type labels; S3: Based on the set multi-level threshold rules, if the anomaly confidence score is within the uncertainty range between high and low thresholds, a collaborative verification mechanism request is initiated to the host node through the LoRa networking unit. S4: The host node coordinates relevant collaborating nodes to perform data synchronization and distributed inference; S5: The host node summarizes the local results of each collaborating node, performs data fusion and final decision-making, triggers corresponding control commands, and reports the decision summary to the management platform.

Claims

1. A predictive maintenance system for equipment based on multimodal data fusion and edge collaborative decision-making, characterized in that: include: On the edge side, the edge side includes at least two smart data acquisition cards. The smart data acquisition cards are deployed on the device for data acquisition and control. The smart data acquisition cards are interconnected through an internal LoRa networking unit to form a dynamic Mesh network. One smart data acquisition card is elected as the host node, and the others are cooperative nodes. The smart data acquisition card corresponding to the host node communicates with the cloud server through a network unit. A unified management platform, comprising a cloud server and a user terminal, wherein the cloud server interacts with the host node for data exchange, and the user terminal accesses the cloud server via the Internet for monitoring and management; and provides AI model management functions, supporting the remote distribution of trained MSAD algorithm model files in the form of differential upgrade packages to designated intelligent data acquisition cards or device groups. The specific process for differential upgrade is as follows: 1) Differential Package Generation: The cloud server compares the new version of the MSAD algorithm model file with the old version of the model file byte by byte, extracts the difference (Delta) to generate a differential upgrade package, and calculates its MD5 value; 2) Remote distribution: The cloud server splits the differential packet into data blocks and distributes them one by one to the designated target intelligent data acquisition card through the host node; In this case, a mesh network is used, and each device in the network has its own number (operated by the wireless module). Users can also match the number with the corresponding deployment location themselves, and upgrade according to the number when upgrade is needed; 3) Verification and Deployment: After the target acquisition card receives the data, it calculates the MD5 value of the received data and compares it with the original MD5 value sent from the cloud to ensure data integrity. After the verification is successful, the acquisition card uses the built-in Bootloader program to merge the differential package with the local old model, generate a new model, and load and run it. 4) Secure rollback: If the new model fails to load or causes system instability, the acquisition card will automatically switch back to the old version and report the failure log to the management platform; The intelligent data acquisition card performs real-time analysis on locally acquired multimodal sensor data based on the built-in MSAD algorithm model and outputs a quantified anomaly confidence score. When the anomaly confidence score is within the uncertainty range between the first high threshold and the second low threshold preset by the system, a collaborative verification mechanism is triggered. The collaborative verification mechanism is jointly completed by the requesting node, the dynamically elected host node, and one or more collaborative nodes, which synchronously acquire data, perform distributed reasoning and result fusion, and finally form a joint decision at the edge.

2. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 1, characterized in that: The intelligent data acquisition card includes a data acquisition unit, a control unit, a core processing unit, a communication unit, and an auxiliary unit. The data acquisition unit, control unit, communication unit, and auxiliary unit are all bidirectionally connected to the core processing unit. The data acquisition unit includes an analog acquisition unit and a digital acquisition unit, and the control unit includes an analog control unit and a digital control unit; The core processing unit is based on the analysis and calculation unit, and the MSAD algorithm model is located within the analysis and calculation unit. The communication unit includes a networking unit and a network unit; the collaboration between the analysis and computing unit and the networking unit jointly supports the MSAD algorithm model and the dynamic collaboration mechanism. The auxiliary unit includes a power supply unit and an indicator unit.

3. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 2, characterized in that: The analysis and computing unit adopts a heterogeneous computing architecture, which includes a multi-core CPU and a dedicated AI acceleration chip for running the Multimodal Perception Anomaly Detection (MSAD) algorithm model.

4. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 2, characterized in that: The networking unit is used to realize inter-node networking communication and uplink data transmission to the master node. The communication methods used for inter-node networking communication include, but are not limited to, LoRa wireless communication modules, and can also use Wi-Fi Mesh, Zigbee, Bluetooth Mesh or other communication technologies that support self-organizing networks; the uplink data transmission methods to the master node include, but are not limited to, wired Ethernet, wireless Wi-Fi, and cellular networks.

5. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 1, characterized in that: The MSAD algorithm model adopts a dual-branch deep neural network structure, including a numerical anomaly perception branch for capturing local feature patterns and a sequence anomaly perception branch for capturing long-term sequence dependencies. The numerical anomaly perception branch adopts a one-dimensional convolutional neural network, and the sequence anomaly perception branch adopts a gated recurrent unit. The MSAD algorithm model performs fusion analysis on the collected multimodal time series data, outputs anomaly confidence scores and fault type labels, and triggers different levels of response according to preset multi-level threshold rules.

6. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 5, characterized in that: The multi-level threshold rule is specifically as follows: If the abnormal confidence score is greater than the first high threshold, local control will be triggered immediately and a high-risk alarm will be reported. If the abnormal confidence score falls within the "uncertainty interval" between the first high threshold and the second low threshold, the collaborative verification mechanism is activated to request verification from neighboring nodes. If the anomaly confidence score is lower than the second lowest threshold, only the data is recorded and no active response is triggered. Multi-level threshold calibration methods require calibration based on historical data and are not fixed. The system provides a self-learning threshold calibration function, as detailed below: 1): Collect historical data under normal equipment conditions and input it into the MSAD model to obtain the confidence distribution. Set the 95th percentile as the low threshold T_low. 2): Collect known fault injection or historical fault data to obtain the confidence distribution, and set the 5th percentile as the high threshold T_high.

7. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 6, characterized in that: The data processing flow of the MSAD algorithm model is as follows: 1) Data preprocessing and alignment: After receiving the raw data collected by the data acquisition unit, a joint denoising algorithm based on wavelet transform is used to filter the high-noise industrial data. The 'db4' wavelet basis function is used to decompose the raw signal into 5 levels. A soft threshold function is used to denoise the high-frequency detail coefficients of each level. After denoising, a dynamic time warping algorithm is used to align the time-series asynchrony between multimodal sensor data. Specifically, during alignment, the curved window size is set to 10, and the distance measurement method is Euclidean distance to ensure the synchronization of different modal data on the time axis. 2): Multi-scale feature extraction, which utilizes the parallel computing capabilities of the AI ​​acceleration chip in the analysis and computing unit to simultaneously extract multi-scale feature values ​​of time domain features, frequency domain features, and time-frequency domain features from the preprocessed data. The system then fuses the extracted feature values ​​into a single feature vector. 3) Anomaly detection and inference: The fused single feature vector is input into the MSAD algorithm model for inference. Specifically, a one-dimensional convolutional neural network of the numerical anomaly perception branch is used to capture local feature patterns, and a gated recurrent unit of the sequence anomaly perception branch is used to capture long-term sequence dependencies. The outputs of the numerical anomaly perception branch and the sequence anomaly perception branch are summarized in the decision fusion layer to form a fused comprehensive feature. The comprehensive feature is fed into the fully connected layer to finally generate a comprehensive anomaly confidence score of 0-1 and a preliminary fault type label. This anomaly confidence score is the direct basis for triggering different levels of response and collaborative verification. 4): Adaptive collaborative decision-making mechanism: The system performs different operations through adaptive decision-making based on preset multi-level threshold rules.

8. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 1, characterized in that: The intelligent data acquisition card uses a resource-aware distributed host election algorithm to elect a host node. During the host node election process, the resource status of each node is communicated through heartbeat packets, and the host node is elected based on a dynamic scoring function. The host node then coordinates member nodes to perform data sharing and distributed collaborative computing, as detailed below: 1) Initial Network Setup and Status Broadcast: After powering on, each node enters network maintenance mode and broadcasts a "heartbeat" packet containing its own device ID, computing power (CPU utilization, memory remaining), power status, and signal strength (RSSI) every T seconds (T is set to 30 seconds by default, and the period can be configured remotely through the management platform). 2) Dynamic host node election: After a node receives a heartbeat packet from another node, it will not immediately trigger a re-election to avoid network jitter; Each node maintains an election timer (the default duration is 3 times the heartbeat cycle, i.e., 90 seconds). When the timer expires, the node collects the latest heartbeat information of all neighboring nodes received during the window period, substitutes it into the scoring function below to calculate the score, and the node with the highest score is identified as the host node. If no host node is specified, all nodes follow this rule: the node with the highest score becomes the host node; other nodes automatically join the network as members. Since CPU_Free (percentage), Memory_Free (MB), Battery_Level (percentage) and RSSI (dBm, usually negative) have different dimensions, they need to be normalized by Min-Max before being substituted into the scoring function to uniformly map them to the [0,1] interval; Specific formula: in, Let these be the minimum and maximum values ​​of the physical quantity in the network within the most recent time window; taking RSSI as an example, let the set of all original RSSI values ​​(negative values, such as -45, -62, -78...) stored in the current sliding window be . The absolute value of all values ​​in the set is obtained. Find the maximum value in. and minimum value Then, the absolute value of RSSI of the current node's heartbeat is normalized using the Min-Max normalization method. Other physical quantities are normalized in the same way based on their respective sliding windows. Substitute the normalized parameters into the following formula: Wherein, CPU_Free is the CPU idle rate, Memory_Free is the remaining memory, Battery_Level is the battery percentage, and RSSI is the signal strength; α, β, γ, and δ are weighting coefficients used to balance the importance of each indicator, satisfying α + β + γ + δ = 1. The system provides a default set of equalization coefficients: α = 0.3, β = 0.2, γ = 0.3, and δ = 0.

2. The node with the highest score becomes the host node; other nodes automatically join the network as members. This election mechanism ensures that the network coordinator is always the node with the best current resources. 3) When a member node's abnormal MSAD confidence score, output by the MSAD algorithm model, falls into the "uncertainty interval," it initiates a "cooperative verification request" to the host node. Upon receiving the "cooperative verification request," the host node, based on the request type and network resource status, instructs one or more other nodes that are complementary in physical location or data type. At this point, these designated nodes become "cooperative nodes," synchronously collecting data and performing local analysis or distributed inference. Each cooperative node feeds back the analysis results to the host node, which then performs information fusion and makes the final decision.

9. The equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making according to claim 8, characterized in that: The criteria for determining "complementarity" include: Spatial complementarity: The physical straight-line distance between the collaborating node and the verified node is less than a preset threshold; Modal complementarity: The sensor type of the cooperating node is different from that of the verified node. For example, the vibration node requests the temperature node and the current node to participate in the verification. The host node selects the node according to the pre-configured cooperative strategy library. The cooperative verification mechanism selects 2 cooperative nodes by default to balance network overhead while ensuring the reliability of the decision. Users can set the upper limit of the number of participating cooperative nodes in the cooperative strategy configuration interface of the management platform. The default is no more than 3.

10. A predictive method for a predictive maintenance system for equipment based on multimodal data fusion and edge collaborative decision-making, characterized in that: The prediction method is based on the equipment predictive maintenance system based on multimodal data fusion and edge collaborative decision-making as described in any one of claims 1 to 9, and its specific prediction method is as follows: S1): The intelligent data acquisition card collects multimodal physical signals from industrial equipment; S2): The analysis and operation unit in the core processing unit uses the MSAD algorithm model to perform real-time analysis on the multimodal physical signals collected in S1). The decision fusion layer fuses the data to obtain a quantified anomaly confidence score and fault type label. S3): Based on the set multi-level threshold rules, if the abnormal confidence score is in the uncertainty range between the high and low thresholds, a collaborative verification mechanism request is initiated to the host node through the LoRa networking unit; S4): The host node coordinates relevant collaborating nodes to perform data synchronization and distributed inference; S5): The host node summarizes the local results of each cooperating node, performs data fusion and final decision-making, triggers corresponding control commands, and reports the decision summary to the management platform.

Images