Equipment fault prediction and health management method and system based on multi-source data fusion

Abstract

The invention discloses an equipment fault prediction and health management method and system based on multi-source data fusion, and relates to the field of system predictive maintenance based on the industrial Internet, and the method comprises the steps: S1, in an equipment health operation stage, synchronously collecting parameters related to equipment operation in multiple dimensions as historical data, a training data set is formed through historical data; s2, on the basis of the training data set, constructing a reference model through a multi-manifold soft contrast learning algorithm; s3, in the real-time operation process of the equipment, keeping the dimension and the frequency consistent with those of the historical data to synchronously collect the real-time operation data flow so as to obtain a real-time sample sequence xt; s4, inputting xt into the reference model to obtain a comprehensive health state dissimilarity value St; and S5, the St is compared with a preset multi-level early warning threshold value, and a corresponding fault processing mode is triggered according to a comparison result. According to the method, normal working condition switching can be prevented from being misjudged as a fault, and the system reliability is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance of systems based on the Industrial Internet. More specifically, this invention relates to a method and system for equipment fault prediction and health management based on multi-source data fusion for key equipment (such as stacker cranes, AGVs, and conveyor lines) in automated warehousing and logistics systems. Background Technology

[0002] Automated warehousing and logistics systems are a core component of modern intelligent manufacturing, and the continuous and stable operation of their core equipment (such as stacker cranes, AGVs, and conveyor lines) is crucial.

[0003] Currently, the maintenance of core equipment in automated warehousing and logistics systems mainly adopts the following two modes: The first is post-failure repair, which involves stopping the machine for repairs after a malfunction. The problem with this is that it can lead to production interruptions and cause huge economic losses.

[0004] The second is regular preventative maintenance, which involves maintenance based on fixed times or operating cycles, regardless of the actual condition of the equipment. This model has two major drawbacks: first, insufficient maintenance, where equipment may malfunction before the maintenance cycle is due; and second, over-maintenance, where equipment is disassembled and replaced even when it is in good condition, wasting manpower and spare parts resources.

[0005] Of course, existing technologies have also introduced sensor threshold-based monitoring, such as monitoring motor temperature and triggering an alarm when it exceeds a certain fixed threshold; however, this method is essentially still a lagging and rudimentary diagnosis rather than a prediction, meaning it cannot identify the trend of equipment performance degradation and can only trigger an alarm when a fault is about to occur or has already occurred, leaving maintenance personnel with very little response time.

[0006] Furthermore, existing sensor threshold monitoring technologies primarily employ single-parameter, single-point monitoring, lacking comprehensive analysis and fusion judgment of the equipment's multi-dimensional operating status (such as vibration, temperature, current, noise, and control parameters). For example, early bearing wear in a motor may only manifest as a slight change in the vibration spectrum, without a temperature increase; temperature monitoring alone cannot detect such potential faults.

[0007] Therefore, it can be seen that the existing maintenance model is lagging behind, passive, and has inaccurate single-parameter monitoring. The main problems are: inability to predict early hidden faults; inability to quantify the decline trend of equipment health status; inability to provide decision-making basis for precise maintenance; and unplanned downtime and resource waste caused by improper maintenance. In other words, the existing technology cannot achieve early, accurate, and forward-looking fault prediction of warehousing and logistics equipment, and it is difficult to support the strategic transformation from "preventive maintenance" to "predictive maintenance". Summary of the Invention

[0008] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0009] To achieve these objectives and other advantages of the present invention, a method for equipment fault prediction and health management based on multi-source data fusion is provided, comprising: S1. During the equipment health operation phase, collect parameters related to equipment operation from multiple dimensions simultaneously as a historical health operation dataset. D train and through historical health operation datasets D train Construct the training dataset; S2. Based on the training dataset, using a multi-manifold soft contrast learning algorithm, from... D train In this process, a benchmark model capable of characterizing various health conditions of the equipment will be constructed. S3. During real-time operation of the equipment, synchronously collect the real-time data stream at the same dimensions and frequency as historical data to obtain a real-time sample sequence. x t ; S4, will x t Input the baseline model to obtain a comprehensive health status anomaly value. S t ; S5. Measure the anomaly value of health status. S t It compares the data with pre-set multi-level warning thresholds to trigger the corresponding fault handling mode based on the comparison result; The benchmark model is constructed in the following ways: S20. Using the Gaussian mixture model unsupervised clustering method, the training dataset is initially divided into K submanifolds, each submanifold corresponding to a stable operating condition. S21. Construct sample pairs based on each manifold and design a deep encoder that maps the sample pairs to a low-dimensional discriminative space. And a soft contrastive learning algorithm is used to optimize the data for each manifold, among which, The subscript represents the depth encoder function itself. θ This represents the set of all weights and bias parameters in the corresponding network. Indicates the input space, Indicates the output space.

[0010] Preferably, in S20, the first... k The data distribution of a manifold is represented as follows: In the above formula,x It is d A dimensionless observation data vector, and , It is the first k Model parameters of each submanifold J It is the number of Gaussian mixture components within the current manifold. It is the first j The mixing weights of Gaussian components, To represent a mean is The covariance matrix is The multivariate Gaussian distribution, p ( ) represents the probability density function, indicating the parameters of the given k-th healthy submanifold. At that time, the probability density of the multidimensional data point x was observed.

[0011] Preferably, a multivariate Gaussian distribution is used. It is characterized by the following formula: In the above formula, Indicates the first k In a healthy submanifold, the first... j Covariance matrix of Gaussian components The inverse matrix, Representing data points x To the corresponding Gaussian component center The square of the Mahalanobis distance.

[0012] Preferably, in S21, the loss function of the soft contrastive learning algorithm... It is characterized by the following formula: In the above formula, It is a set of positive sample pairs consisting of data-augmented samples within the same submanifold. Anchor point x i The negative sample set, s() is the cosine similarity. For temperature hyperparameters, For positive sample pairs from the same healthy submanifold, These are soft negative sample pairs from different but adjacent healthy submanifolds.

[0013] Preferably, in S1, the historical data includes vibration signals, temperature, operating current, noise signals, and servo drive parameters output by the device controller.

[0014] Preferably, in S4, the health status anomaly value S t The method of obtaining it is: S40, The encoder processes the real-time sample sequence using the following formula. x t After processing, the corresponding representation vector is obtained. Z t : S41. Calculated using the following formula Z t The weighted minimum distance to all healthy manifold centers is used as the health state anomaly value: in, for Z t To the k manifold center The Mahalanobis distance or Euclidean distance. For the first k Weights of health confidence in historical data for each manifold.

[0015] Preferably, in S5, the multi-level warning threshold includes: attention threshold, warning threshold, and danger threshold; In S5, the processing mode is as follows: When the health status is abnormal S t When the threshold is continuously exceeded and shows an upward trend, an early warning is triggered, and an early warning diagnostic report is created in the database; When the health status is abnormal S t When the warning threshold is approached, a maintenance work order is triggered, and an emergency maintenance work order containing the fault mode and affected components is generated. The affected components are located using the following formula: in, For the corresponding equipment component j 3D sensing signal, To obtain the partial derivative.

[0016] A fault prediction and health management system includes: a multi-source data acquisition layer, a health model construction layer, a real-time status assessment layer, and an intelligent early warning and decision-making layer; The data acquisition layer includes a multi-source data acquisition module deployed on the device and a PLC controller that communicates with it. The data acquisition module includes a vibration sensor, a temperature sensor, a current transformer, and a noise microphone. The health model construction layer, real-time status assessment layer, and intelligent early warning and decision-making layer are all deployed on the server, and the multi-source data acquisition layer communicates with the server via industrial Ethernet.

[0017] The present invention has at least the following beneficial effects: Firstly, this invention can avoid misjudging normal operating condition switching as a fault, significantly improving system reliability; that is, this invention can greatly reduce the false alarm rate.

[0018] Secondly, this invention can detect systemic minor deviations in the early stages of a fault (such as slight wear of components), enabling earlier warnings. In other words, this invention can improve the lead time for warnings.

[0019] Thirdly, when generating an early warning system, this invention can indicate faulty components. For example, the emergency maintenance work order can include "suspected abnormal bearing vibration under high-speed conditions," providing maintenance personnel with precise inspection directions and achieving a leap from "monitoring" to "diagnosis."

[0020] Fourth, the present invention can safely incorporate new health data into existing manifolds or form new manifolds through online learning strategies, that is, the present invention can achieve scalability of manifold structures.

[0021] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the module composition and data flow of the fault prediction and health management system in this invention; Figure 2 This is the overall flowchart of the fault prediction and health management method in this invention; Figure 3 A comparison chart of the health benchmark model and real-time data in this invention; Figure 4 This is a trend graph showing the change of "health status anomaly value" over time in the invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0024] This invention provides a complete fault prediction and health management system integrating "data perception, model building, status assessment, and prediction and early warning," including: Multi-source data acquisition layer: used to acquire multi-dimensional operating parameters of the equipment in real time, including but not limited to vibration signals, temperature, operating current, noise signals, and servo drive parameters output by the equipment controller (such as following error and torque output).

[0025] Health Model Construction Layer: This layer collects historical data of the multidimensional operating parameters during fault-free equipment operation to construct a multidimensional benchmark model characterizing the equipment's health baseline state. This model is not based on fixed thresholds but rather on a dynamic cluster of normal operating condition data.

[0026] Real-time status assessment layer: This layer compares the multi-dimensional operating parameters collected in real time with the multi-dimensional benchmark model and calculates a comprehensive "health status anomaly value" through a specific fusion algorithm. This value quantifies the degree of deviation between the current status and the health benchmark.

[0027] Intelligent early warning and decision-making layer: It is used to generate early warning information of different levels based on the changing trend of the "health status anomaly value" and the preset early warning threshold, and automatically output maintenance diagnosis suggestions.

[0028] Furthermore, the present invention also provides a method for equipment fault prediction and health management based on multi-source data fusion, which mainly includes the following steps: S1. During the equipment health operation phase, a historical health operation dataset Dtrain is synchronously collected, containing multi-dimensional operating parameters, to form the training dataset. Each sample... It is d A dimension vector represents the device's state at a specific point in time, based on historical health data collected during a particular phase. The dimension encompasses key physical parameters such as triaxial vibration, current, and temperature.

[0029] S2. Based on the training dataset, using a multi-manifold soft contrast learning algorithm, from... D train We construct a benchmark model capable of accurately depicting various health conditions of the equipment. This process involves two key steps: 1. Selecting a healthy manifold: Using Gaussian mixture model unsupervised clustering, the health data is initially divided into... k Each manifold corresponds to a steady-state condition (such as high speed or full load). k The data distribution of a manifold can be represented as: in, p ( ) represents the probability density function. Denotes the parameters of a given k-th healthy submanifold. When the probability density of the observed multidimensional data point x is... It is a d-dimensional observation data vector. p Let represent the probability density function, given the th k Parameters of a healthy submanifold At that time, multidimensional data points were observed. xThe probability density, It is the first k Model parameters of each submanifold J It is the number of Gaussian mixture components within the current manifold. It is the first j Mixing weights of Gaussian components To represent a mean is The covariance matrix is The multivariate Gaussian distribution, and In the above formula, Indicates the first k In a healthy submanifold, the first... Covariance matrix of Gaussian components The inverse matrix, It is a key component for calculating Mahalanobis distance, and its expression is... The calculation is for data points. To the corresponding Gaussian component center The square of the Mahalanobis distance.

[0030] 2. Adaptive Representation Learning: Design a Deep Encoder This involves mapping the data to a low-dimensional discriminative space, where sample representations within the same healthy submanifold should be close to each other, while sample representations between different submanifolds should maintain a "reasonable" distance. This represents the deep encoder function itself, which is a mapping function implemented by a neural network. The subscript... θ This represents the set of all weights and bias parameters in the corresponding network. Represents the input space (representing a d-dimensional real vector space). The output space, also known as the representation space, represents an m-dimensional real vector space. Optimization is achieved using soft contrastive learning loss, and its core innovation lies in constructing sample pairs: Positive sample pairs : From the same healthy submanifold.

[0031] Soft negative sample pairs : Healthy submanifolds that are different but adjacent.

[0032] The soft contrastive learning loss function is defined as: in, It is a set of positive sample pairs, consisting of data-augmented samples within the same submanifold; Anchor point x iThe key innovation of this negative sample set lies in including samples from other healthy submanifolds as "soft negative samples". s() represents the cosine similarity. This refers to temperature hyperparameters.

[0033] The core of this loss function is that it does not force all healthy samples to be compressed to the same point, but allows different health conditions (manifolds) to be appropriately separated in the representation space, thereby more accurately describing the complex health state space.

[0034] After training, the health baseline model consists of two parts: the encoder and the center of the health manifold in the space. Together, they constitute a "multidimensional baseline map" of the device's health status.

[0035] S3. During the real-time operation of the equipment, synchronously collect real-time operation data streams at the same dimensions and frequency as historical data to form a real-time sample sequence. .

[0036] S4, such as Figure 3 As shown, the real-time sample sequence x t Inputting the aforementioned health benchmark model, a comprehensive "health status anomaly value" is calculated. First, its representation vector is obtained through an encoder. Then, calculations were performed. Z t The weighted minimum distance to all healthy manifold centers is used as a comprehensive "health status anomaly value". S t : in, To reach the first k manifold center The Mahalanobis distance or Euclidean distance. The weights are based on the health confidence scores of the corresponding manifold historical data. S t The larger the value, the more severe the deviation of the current state from all known healthy patterns.

[0037] S5: Set multi-level early warning thresholds (e.g.) Figure 4 The system displays three thresholds: attention threshold, warning threshold, and danger threshold. When the abnormality value continuously exceeds the "attention threshold" and shows an upward trend, an early warning is triggered. When it approaches the "warning threshold," a maintenance work order is triggered, and a warning diagnostic report is generated, indicating potential fault modes and affected components. The system locates the most likely faulty component based on the following formula: in, For the first jDimensional sensor signals (corresponding to specific components). The dimension with the largest partial derivative indicates the main contributing source of the anomaly, thereby generating an early warning diagnostic report containing potential failure modes and affected components.

[0038] Example 1: In this embodiment, the architecture of the obstacle prediction and health management system is as follows: Figure 1 As shown, it mainly includes: Data acquisition layer (101): Composed of vibration sensors, temperature sensors, current transformers, noise microphones, etc. deployed on the equipment, and communicates with the equipment PLC controller through an IoT gateway to obtain servo drive parameters. All data is transmitted to the edge server or central server via industrial Ethernet.

[0039] Health model building layer (102): Deployed on the server, responsible for storing historical data and running health benchmark model algorithms.

[0040] Real-time status assessment layer (103): Performs comparison calculations between real-time data and the model to obtain the "health status anomaly value".

[0041] Intelligent early warning and decision-making layer (104): Provides a human-computer interaction interface to display the real-time health status of the equipment, the curve of abnormal value change, the list of early warning information and the generated early warning diagnosis form.

[0042] Example 2: like Figure 2 As shown, this example mainly uses the stacker crane lifting motor as an example to illustrate the execution steps of the fault prediction and health management system: S201: Start. After the stacker crane is installed and commissioned, it should run stably for at least one month. During this period, multi-source data (vibration (3-axis acceleration), bearing temperature, three-phase current, and drive torque output value) of its lifting motor should be collected as health baseline data.

[0043] S202: Construction of the health baseline model. A multi-manifold soft computing algorithm is used to train the health baseline data, forming a model that encompasses the vast majority of normal data points. This model serves as the "health fingerprint" of the motor.

[0044] S203: Real-time data synchronization acquisition. During the daily operation of the motor, its vibration (3-axis acceleration), bearing temperature, three-phase current, and driver torque output value are continuously collected.

[0045] S204: Health Status Assessment. The real-time data vector collected in S203 is input into the multi-manifold soft computing model constructed in S202. The Mahalanobis distance from the data vector to the model center is calculated, and this distance is normalized and used as the "health status anomaly value" for this sampling.

[0046] S205: Trend Analysis and Early Warning Judgment. The system does not rely solely on a single abnormal value for judgment, but continuously records and analyzes the moving average trend of that value.

[0047] If the anomaly value trend is stable and below the first-level threshold, return to S203 to continue monitoring.

[0048] If the anomaly value exceeds the first-level threshold (note the threshold) for 10 consecutive sampling periods and the slope shows an upward trend, then execute S206.

[0049] S206: Generate early warning and diagnostic report. The system triggers an "early warning" and creates an early warning diagnostic report in the database. The report may include: "Warning object: No. 1 stacker crane lifting motor; Warning time: [specific time]; Current anomaly value: 0.15; Trend: Upward; Main contributing parameter: Vibration high-frequency energy increase of 35%; Suspected fault mode: Early bearing wear; Recommended measures: Focus on checking the bearing during the next maintenance and increase the monitoring frequency of this point." S207: When the abnormality value rises further and exceeds the secondary threshold (warning threshold), the system automatically creates an emergency maintenance work order in the enterprise's computerized maintenance management system and notifies the maintenance team.

[0050] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0051] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for equipment fault prediction and health management based on multi-source data fusion, characterized in that, include: S1. During the equipment health operation phase, collect parameters related to equipment operation from multiple dimensions simultaneously as a historical health operation dataset. D train and through historical health operation datasets D train Construct the training dataset; S2. Based on the training dataset, using a multi-manifold soft contrast learning algorithm, from... D train In this process, a benchmark model capable of characterizing various health conditions of the equipment will be constructed. S3. During real-time operation of the equipment, synchronously collect the real-time data stream at the same dimensions and frequency as historical data to obtain a real-time sample sequence. x t ; S4, will x t Input the baseline model to obtain a comprehensive health status anomaly value. S t ; S5. Measure the anomaly value of health status. S t It compares the data with pre-set multi-level warning thresholds to trigger the corresponding fault handling mode based on the comparison result; The benchmark model is constructed in the following ways: S20. Using the Gaussian mixture model unsupervised clustering method, the training dataset is initially divided into K submanifolds, each submanifold corresponding to a stable operating condition. S21. Construct sample pairs based on each manifold and design a deep encoder that maps the sample pairs to a low-dimensional discriminative space. And a soft contrastive learning algorithm is used to optimize the data for each manifold, among which, The subscript represents the depth encoder function itself. θ This represents the set of all weights and bias parameters in the corresponding network. Indicates the input space, Indicates the output space.

2. The equipment fault prediction and health management method based on multi-source data fusion as described in claim 1, characterized in that, In S20, the data distribution of the k-th manifold is represented as: In the above formula, x It is a d-dimensional observation data vector, and , It is the first k Model parameters of each submanifold J It is the number of Gaussian mixture components within the current manifold. It is the first j The mixing weights of Gaussian components, Represents a mean of The covariance matrix is The multivariate Gaussian distribution, p ( ) represents the probability density function. Denotes the parameters of a given k-th healthy submanifold. At that time, the probability density of the multidimensional data point x was observed.

3. The equipment fault prediction and health management method based on multi-source data fusion as described in claim 1, characterized in that, Multivariate Gaussian distribution It is characterized by the following formula: In the above formula, Let the covariance matrix of the j-th Gaussian component in the k-th healthy submanifold be denoted as . The inverse matrix, This represents the distance from data point x to the corresponding Gaussian component center. The square of the Mahalanobis distance.

4. The equipment fault prediction and health management method based on multi-source data fusion as described in claim 1, characterized in that, In S21, the loss function of the soft contrastive learning algorithm It is characterized by the following formula: In the above formula, It is a set of positive sample pairs consisting of data-augmented samples within the same submanifold. Anchor point x i The negative sample set, s() is the cosine similarity. For temperature hyperparameters, For positive sample pairs from the same healthy submanifold, These are soft negative sample pairs from different but adjacent healthy submanifolds.

5. The equipment fault prediction and health management method based on multi-source data fusion as described in claim 1, characterized in that, In S1, the historical data includes vibration signals, temperature, operating current, noise signals, and servo drive parameters output by the device controller.

6. The equipment fault prediction and health management method based on multi-source data fusion as described in claim 1, characterized in that, In S4, the health status anomaly value S t The method of obtaining it is: S40, The encoder processes the real-time sample sequence using the following formula. x t After processing, the corresponding representation vector is obtained. Z t : S41. Calculated using the following formula Z t The weighted minimum distance to all healthy manifold centers is used as the health state anomaly value: in, for Z t To the k manifold center The Mahalanobis distance or Euclidean distance. For the first k Weights of health confidence in historical data for each manifold.

7. The equipment fault prediction and health management method based on multi-source data fusion as described in claim 1, characterized in that, In S5, the multi-level warning thresholds include: attention threshold, warning threshold, and danger threshold; In S5, the processing mode is as follows: When the health status is abnormal S t When the threshold is continuously exceeded and shows an upward trend, an early warning is triggered, and an early warning diagnostic report is created in the database; When the health status is abnormal S t When the warning threshold is approached, a maintenance work order is triggered, and an emergency maintenance work order containing the fault mode and affected components is generated. The affected components are located using the following formula: in, For the corresponding equipment component j 3D sensing signal, To obtain the partial derivative.

8. A fault prediction and health management system, applied in the equipment fault prediction and health management method based on multi-source data fusion as described in any one of claims 1-7, characterized in that, include: Multi-source data acquisition layer, health model construction layer, real-time status assessment layer, intelligent early warning and decision-making layer; The multi-source data acquisition layer includes a data acquisition module deployed on the device and a PLC controller that communicates with it. The data acquisition module includes a vibration sensor, a temperature sensor, a current transformer, and a noise microphone. The health model construction layer, real-time status assessment layer, and intelligent early warning and decision-making layer are all deployed on the server, and the multi-source data acquisition layer communicates with the server via industrial Ethernet.

Images