Adaptive methods, systems, equipment, and media for detecting anomalies in industrial production lines at different process stages.

Abstract

This application provides an adaptive industrial production line anomaly detection method, system, equipment, and medium for process stages. The method acquires the current process stage of the production line, along with its execution sequence, elapsed runtime, remaining runtime, and multi-source heterogeneous data. It then determines a first set of anomaly detection models and a first pre-configured weight combination for these models. The method determines the operating state of the current process stage. If the current process stage is in a stable state, an anomaly score is assigned. If the current process stage is in a transitional state, the anomaly score is determined. If the anomaly score exceeds a dynamic scoring threshold, the inference conditions are matched with pre-built process diagnostic rules or fault trees to determine the cause of the anomaly in the current process stage. This application enables dynamic switching of the detection model combination with the process stage, ensuring the continuity and stability of detection during the transition phase.

Description

Technical Field

[0001] This application relates to the fields of industrial automation and intelligent manufacturing technology, and in particular to methods, systems, equipment and media for adaptive detection of anomalies in industrial production lines at different process stages. Background Technology

[0002] Industrial production lines (e.g., welding, assembly, machining) typically comprise multiple process stages (e.g., loading, welding, inspection, unloading, etc.), with significant differences in data distribution (e.g., mean, variance, spectral features, etc.) across different stages. Existing anomaly detection methods for process stages include: fixed threshold / fixed model methods, single model methods, dual-domain feature fusion methods, and unsupervised anomaly detection methods based on adaptive deep representation learning. Fixed threshold / fixed model methods cannot adapt to data drift between stages, generating numerous false positives or false negatives during stage transitions. Single model methods, such as Isolation Forest, are insensitive to temporal dependency anomalies, while LSTM (Long Short-Term Memory) is not. Long Short-Term Memory (LSTM) networks are slow to respond to point anomalies, making it difficult to cover multiple anomaly types. Dual-domain feature fusion methods only fuse time-frequency domain features at the feature layer, essentially remaining a single model architecture with poor stage adaptability. Unsupervised anomaly detection methods based on adaptive deep representation learning only involve dynamic threshold adjustment and do not achieve stage adaptation. In addition, some technologies use hybrid expert models, but their selection is based on data features rather than explicit process stage semantics, resulting in uninterpretable routing decisions and difficulty in ensuring detection stability during stage switching. Fault diagnosis methods based on model routing mechanisms are based on equipment type (static division) rather than process stage (dynamic change), and are a "one-to-one" selection method between equipment type and model, without involving multi-model combination and weight configuration.

[0003] Therefore, there is an urgent need for a parallel fusion model and an adaptive industrial production line anomaly detection method for process stages. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, this application provides an adaptive industrial production line anomaly detection method, system, equipment and medium for process stages to solve the above-mentioned technical problems.

[0005] According to one aspect of the embodiments of this application, an adaptive industrial production line anomaly detection method is provided, comprising: acquiring the current process stage of the production line, and the execution order, elapsed runtime, remaining runtime, and multi-source heterogeneous data of the current process stage; the multi-source heterogeneous data of the current process stage includes: vibration data, current data, and voltage data; determining a first anomaly detection model set and a first pre-configured weight combination of the first anomaly detection model set according to the current process stage; determining the operating state of the current process stage according to the execution order, the elapsed runtime, and the remaining runtime; the operating state includes: a stable state and a transitional state; if the current process stage is in a stable state, determining an anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination; if the current process stage is in a transitional state, determining an anomaly score of the current process stage; if the current process stage is in a transitional state, determining an anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination; if the current process stage is in a transitional state, determining an anomaly score of the current process stage based on .... Based on the adjacent process stages of the current process stage, a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models are determined. An anomaly score for the current process stage is determined based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination. If the anomaly score is greater than a dynamic scoring threshold, the current process stage is determined to be in an abnormal state. The current process stage, the equipment source of the multi-source heterogeneous data of the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data of the current process stage are used as inference conditions and matched with pre-built process diagnostic rules or fault trees to obtain the reason for the abnormal state triggered by the current process stage. The dynamic scoring threshold is determined based on the feature data of the multi-source heterogeneous data within a preset number of sliding windows.

[0006] According to another aspect of the embodiments of this application, an industrial production line anomaly detection system with process stage adaptation is also provided, comprising: a data acquisition module, configured to acquire the current process stage of the production line, and the execution order, elapsed runtime, remaining runtime, and multi-source heterogeneous data of the current process stage; the multi-source heterogeneous data of the current process stage includes: vibration data, current data, and voltage data; a state determination module, configured to determine a first anomaly detection model set and a first pre-configured weight combination of the first anomaly detection model set according to the current process stage; and determine the operating state of the current process stage according to the execution order, elapsed runtime, and remaining runtime; the operating state includes: a stable state and a transitional state; a stable state score determination module, configured to determine an anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination if the current process stage is in a stable state; and a transitional state score determination module. The first process stage is used to determine a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models based on the adjacent process stages if the current process stage is in a transitional state; and to determine the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination; the cause reasoning module is used to determine that the current process stage is in an abnormal state if the anomaly score is greater than the dynamic scoring threshold, and then the current process stage, the equipment source of the multi-source heterogeneous data of the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data of the current process stage are used as reasoning conditions, and are matched with pre-built process diagnosis rules or fault trees to obtain the cause of the current process stage triggering the abnormal state; the dynamic scoring threshold is determined based on the feature data of the multi-source heterogeneous data within a preset number of sliding windows.

[0007] According to one aspect of the present invention, an electronic device is provided, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device enables the industrial production line anomaly detection method with process stage adaptation as described above.

[0008] According to one aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a computer processor, causes the computer to perform the aforementioned process-stage adaptive industrial production line anomaly detection method.

[0009] The beneficial effects of this application are as follows: This application obtains the current process stage of the production line, as well as the execution order, runtime, remaining runtime, and multi-source heterogeneous data of the current process stage. Based on the current process stage, it determines a first set of anomaly detection models and a first pre-configured weight combination of the first set of anomaly detection models. Based on the execution order, runtime, and remaining runtime, it determines the operating status of the current process stage. If the current process stage is in a stable state, it determines the anomaly score of the current process stage based on the multi-source heterogeneous data, the first set of anomaly detection models, and the first pre-configured weight combination. If the current process stage is in a transitional state, it determines a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models based on the adjacent process stages. Based on the multi-source heterogeneous data, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, it determines the anomaly score of the current process stage. If the anomaly score is greater than the dynamic scoring threshold, it determines that the current process stage is in a transitional state. If a process stage is in an abnormal state, the current process stage, the equipment source of the multi-source heterogeneous data in the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data in the current process stage are used as inference conditions. These conditions are matched with pre-built process diagnosis rules or fault trees to determine the cause of the abnormal state triggered in the current process stage. The above process, by determining the first set of anomaly detection models and the first pre-configured weight combination of the first set of anomaly detection models based on the current process stage, enables the dynamic switching of the detection model combination with the process stage, avoiding the performance deficiency of a single model in all stages. By distinguishing between stable states and transition stages, the detection blind spots at the moment of switching in the transition stage are eliminated, significantly reducing the false alarm rate and false negative rate, and ensuring the continuity and stability of detection at the moment of switching in the transition stage. In addition, the process of determining the anomaly score of the current process stage and inferring the cause of the abnormal state triggered in the current process stage can provide traceable and verifiable model routing decisions, enhance the interpretability of detection results, and meet the millisecond-level real-time requirements of industrial production lines while ensuring high accuracy.

[0010] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0012] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application;

[0013] Figure 2 This is a flowchart illustrating an exemplary embodiment of the present application of a process-stage adaptive industrial production line anomaly detection method;

[0014] Figure 3 This is a schematic diagram of the architecture of an industrial production line anomaly detection system, as illustrated in an exemplary embodiment of this application.

[0015] Figure 4 This is a diagram illustrating the adaptive collaboration within a multidimensional adaptive optimization module, as shown in an exemplary embodiment of this application.

[0016] Figure 5 This is a schematic diagram illustrating the dynamic change process of model combination and weights before and after stage switching, as shown in an exemplary embodiment of this application.

[0017] Figure 6 This is a block diagram illustrating an adaptive industrial production line anomaly detection system for process stages, as shown in an exemplary embodiment of this application.

[0018] Figure 7 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown. Detailed Implementation

[0019] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0020] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0021] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0022] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application.

[0023] Reference Figure 1 As shown, the system architecture may include a storage device 101 and an electronic device 102. The electronic device 102 may be at least one of a desktop graphics processing unit (GPU) computer, a GPU computing cluster, a neural network computer, etc. Those skilled in the art can use the electronic device 102 to obtain the current process stage of the production line, as well as the execution order, runtime, remaining runtime, and multi-source heterogeneous data of the current process stage. Based on the current process stage, a first set of anomaly detection models and a first pre-configured weight combination of the first anomaly detection model set are determined. Based on the execution order, runtime, and remaining runtime, the operating state of the current process stage is determined. If the current process stage is in a stable state, the anomaly score of the current process stage is determined based on the multi-source heterogeneous data, the first set of anomaly detection models, and the first pre-configured weight combination. If the current process stage is in a transitional state, the anomaly score is determined based on the adjacent process stages. A second set of anomaly detection models and a second pre-configured weight combination for the second set of anomaly detection models are defined. Based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, the anomaly score of the current process stage is determined. If the anomaly score is greater than the dynamic scoring threshold, the current process stage is determined to be in an abnormal state. The current process stage, the equipment source of the multi-source heterogeneous data of the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data of the current process stage are used as inference conditions and matched with pre-built process diagnosis rules or fault trees to obtain the reason for the abnormal state triggered by the current process stage. The storage device 101 is used to store the current process stage of the production line, as well as the execution order, runtime, remaining runtime, and multi-source heterogeneous data of the current process stage, and provides the current process stage of the production line, as well as the execution order, runtime, remaining runtime, and multi-source heterogeneous data of the current process stage, to the electronic device 102 for processing.

[0024] Schematic, after acquiring the current process stage of the production line in the storage device 101, as well as the execution order, elapsed runtime, remaining runtime, and multi-source heterogeneous data of the current process stage, the electronic device 102 determines a first anomaly detection model set and a first pre-configured weight combination of the first anomaly detection model set based on the current process stage. Based on the execution order, elapsed runtime, and remaining runtime, it determines the operating state of the current process stage. If the current process stage is in a stable state, it determines an anomaly score based on the multi-source heterogeneous data, the first anomaly detection model set, and the first pre-configured weight combination of the current process stage. If the current process stage is in a transitional state, it determines a second anomaly detection model set and a second pre-configured weight combination of the second anomaly detection model set based on the adjacent process stages. Based on the multi-source heterogeneous data, the first anomaly detection model set, the first pre-configured weight combination, the second anomaly detection model set, and the second pre-configured weight combination of the current process stage, it determines an anomaly score for the current process stage. If the anomaly score is greater than the dynamic scoring threshold, it is judged that... If the current process stage is determined to be in an abnormal state, then the current process stage, the equipment source of the multi-source heterogeneous data of the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data of the current process stage are used as inference conditions. These conditions are matched with pre-built process diagnosis rules or fault trees to obtain the cause of the abnormal state triggered by the current process stage. The above process, by determining the first set of anomaly detection models and the first pre-configured weight combination of the first set of anomaly detection models based on the current process stage, enables the dynamic switching of the detection model combination with the process stage, avoiding the performance deficiency of a single model throughout the entire stage. By distinguishing between stable state and transition stage, the detection blind spot at the moment of switching in the transition stage is eliminated, significantly reducing the false alarm rate and false negative rate, and ensuring the continuity and stability of detection at the moment of switching in the transition stage. In addition, the process of determining the anomaly score of the current process stage and inferring the cause of the abnormal state triggered by the current process stage can provide traceable and verifiable model routing decisions, enhance the interpretability of detection results, and meet the millisecond-level real-time requirements of industrial production lines while ensuring high accuracy.

[0025] It should be noted that the process-stage adaptive industrial production line anomaly detection method provided in this application embodiment is generally executed by electronic device 102, and correspondingly, the process-stage adaptive industrial production line anomaly detection system is generally set in electronic device 102.

[0026] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0027] Figure 2 This is a flowchart illustrating an exemplary embodiment of an industrial production line anomaly detection method with process stage adaptation, as shown in this application. (Refer to...) Figure 2As shown, the adaptive industrial production line anomaly detection method for this process stage includes at least steps S210 to S250, which are described in detail below:

[0028] In step S210, the current process stage of the production line, as well as the execution order, elapsed runtime, remaining runtime, and multi-source heterogeneous data of the current process stage are obtained. In one embodiment of this application, the multi-source heterogeneous data of the current process stage includes vibration data, current data, voltage data, etc. The current process stage and its execution order are characterized by stage status codes in the status register of the programmable logic controller (PLC) on the industrial production line. For example, 1 represents the loading stage, 2 represents the welding stage, and 3 represents the unloading stage. The current process stage is characterized by stage status codes in the status register of the PLC on the industrial production line, eliminating the need for feature analysis calculations on sensor data, effectively reducing processing latency and keeping the latency within less than 1ms. The elapsed runtime is stored in the PLC after being timed by a timer. The remaining runtime is the difference between the total runtime and the elapsed runtime, and is stored in the PLC after the remaining runtime is calculated. Multi-source heterogeneous data is collected by sensors and stored within the PLC. Vibration data is detected by vibration sensors, current data by current sensors, and voltage data by voltage sensors. Equipment requiring vibration data monitoring consists of key rotating machinery or transmission components in industrial production lines, including drive motors, gearboxes, fluid pumps, industrial fans, machine tool spindles, or welding positioners. Equipment requiring current data monitoring consists of power drive units in industrial production lines, including AC inverters, DC drivers, and servo drivers. Equipment requiring voltage data monitoring consists of core power supply units in industrial production lines, including incoming power distribution cabinet busbars and the output terminals of dedicated power transformers. All multi-source heterogeneous data can be directly read through the PLC's industrial Ethernet interface, ensuring real-time data retrieval.

[0029] In step S220, a first set of anomaly detection models and a first pre-configured weight combination of the first set of anomaly detection models are determined according to the current process stage; the running state of the current process stage is determined according to the execution order, the elapsed runtime, and the remaining runtime. In one embodiment of this application, the running state includes: a stable state, a transitional state, etc. The process of determining the first set of anomaly detection models and the first pre-configured weight combination of the first set of anomaly detection models according to the current process stage includes: after obtaining the current process stage, finding the first set of anomaly detection models corresponding to the current process stage and the pre-configured weights of each anomaly detection model in the first set of anomaly detection models by searching a pre-established "stage-model configuration" mapping table, and the pre-configured weights of each anomaly detection model in the first set of anomaly detection models form the first pre-configured weight combination. The operating states include: stable state, transition state, etc. The above method of determining the first anomaly detection model set and the first pre-configured weight combination of the first anomaly detection model set according to the current process stage can ensure that different process stages match the corresponding first anomaly detection model set and the first pre-configured weight combination of the first anomaly detection model set. This realizes the dynamic switching of the detection model combination with the process stage, avoids the performance deficiency of a single model in all stages, and improves the flexibility and accuracy of anomaly detection.

[0030] In step S230, if the current process stage is in a stable state, the anomaly score of the current process stage is determined based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination. In one embodiment of this application, the process of determining the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination includes: preprocessing the multi-source heterogeneous data of the current process stage to obtain preprocessed multi-source heterogeneous data; inputting the preprocessed multi-source heterogeneous data into each anomaly detection model in the first anomaly detection model set to obtain a first initial anomaly score set; and weighting and fusing the scores in the first initial anomaly score set according to the first pre-configured weight combination to obtain the anomaly score. This process, by comprehensively utilizing multi-source heterogeneous data to obtain a first initial anomaly score set output by multiple models, and weighting and fusing the scores in the first initial anomaly score set output by multiple models to obtain the anomaly score, fully leverages the detection advantages of different anomaly detection models for different anomaly types compared to single data and single model detection methods, further improving the accuracy of anomaly detection.

[0031] In step S240, if the current process stage is in a transition state, a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models are determined based on the adjacent process stages of the current process stage. Then, based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, the anomaly score of the current process stage is determined. In one embodiment of this application, the transition state includes a pre-transition state and a post-filtering state. If the current process stage is in the pre-transition state, a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models are determined based on the previous process stage of the current process stage. Then, based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, the anomaly score of the current process stage is determined, thus completing the smooth transition between the previous process stage and the current process stage. If the current process stage is in a post-transition state, a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models are determined based on the next process stage. Then, based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, the anomaly score of the current process stage is determined. After determining the anomaly score of the current process stage, the anomaly score of the next process stage is determined based on the multi-source heterogeneous data of the next process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, thus completing the smooth transition between the current and next process stages. The process of determining the anomaly score of the next process stage based on the multi-source heterogeneous data of the next process stage is the same as the process of determining the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage.

[0032] In step S250, if the anomaly score is greater than the dynamic scoring threshold, the current process stage is determined to be in an abnormal state. The current process stage, the equipment source of the multi-source heterogeneous data in the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data in the current process stage are used as inference conditions. These are then matched with pre-built process diagnosis rules or fault trees to determine the cause of the abnormal state in the current process stage. In one embodiment of this application, the dynamic scoring threshold is determined based on the feature data of the multi-source heterogeneous data within a preset sliding window. The vibration data features include: time-domain statistical features, frequency-domain features, and time-frequency-domain features. The time-domain statistical features include: root mean square, peak value, skewness, and kurtosis. The frequency-domain features are extracted through Fourier transform and include: energy proportion of each frequency band, centroid frequency, and frequency variance. The time-frequency-domain features are obtained through wavelet decomposition and include: energy statistical values ​​of each wavelet component. The current data features include: average current, current fluctuation variance, and harmonic component amplitude proportion. The voltage data features include: effective voltage value, voltage fluctuation rate, and voltage deviation rate. Pre-built process diagnostic rules or fault trees are stored in an expert knowledge base. Rule matching with the pre-built process diagnostic rules or fault trees is achieved through an inference engine. The inference engine can be selected according to the actual situation, such as a large language inference model or a fault tree modeling and simulation platform for industrial control systems.

[0033] In one embodiment of this application, by determining a first set of anomaly detection models and a first pre-configured weight combination of the first set of anomaly detection models according to the current process stage, the detection model combination is dynamically switched with the process stage, avoiding the performance deficiency of a single model throughout the entire process stage; by distinguishing between stable state and transition stage, the detection blind spot at the moment of switching in the transition stage is eliminated, significantly reducing the false alarm rate and false negative rate, and ensuring the continuity and stability of detection at the moment of switching in the transition stage; in addition, the process of determining the anomaly score of the current process stage and reasoning about the cause of the anomaly state triggered by the current process stage can provide traceable and verifiable model routing decisions, enhance the interpretability of the detection results, and meet the millisecond-level real-time requirements of industrial production lines while ensuring high accuracy.

[0034] In one embodiment of this application, if the transition state includes a pre-transition state and a post-filtering state, the process of determining the operating state of the current process stage based on the execution order, the elapsed runtime, and the remaining runtime includes:

[0035] If the execution order is first, the runtime already elapsed is greater than or equal to the preset cold start state duration, and the remaining runtime is greater than the transition state duration, then the current process stage is determined to be in a stable state. In one embodiment of this application, to reduce the occurrence of transition stages, the anomaly detection model set for the first executed process stage in the cold start state is consistent with the anomaly detection model set for the first executed process stage in the stable state, and the anomaly detection model set for the first executed process stage is set with full consideration of the existence of a cold start state. The total runtime of the first executed process stage is greater than the sum of the preset cold start state duration and the transition state duration, and the preset cold start state duration is set according to the process line and the type of process stage. The transition state duration is determined based on the multi-source heterogeneous data distribution of the current process stage and the similarity between the multi-source heterogeneous data distributions of adjacent process stages. The calculation formula for the transition state duration is as follows:

[0036] Equation (1)

[0037] in, Indicates the duration of the transition state. Indicates the base duration (e.g., 2 windows). This indicates the stretch factor (e.g., 4). Denotes KL divergence, This indicates the distribution of multi-source heterogeneous data at the current process stage. This represents the distribution of multi-source heterogeneous data in adjacent process stages. This indicates the shortest transition half-window (e.g., 2 windows). This represents the longest transition half-window (e.g., 6 windows). Formula (1) is used to characterize that the greater the difference in data distribution, the longer the transition state takes, and the smaller the difference in data distribution, the shorter the transition state takes. It can automatically adapt the length of the transition window according to the actual data characteristics of different process stages, and avoid detection errors caused by fixed windows.

[0038] In one embodiment of this application, the formula for calculating the KL divergence between the multi-source heterogeneous data distribution in the current process stage and the multi-source heterogeneous data distribution in adjacent process stages is as follows:

[0039] Equation (2)

[0040] in, Denotes KL divergence, This indicates the distribution of multi-source heterogeneous data at the current process stage. This represents the distribution of multi-source heterogeneous data in adjacent process stages. This indicates the status of multi-source heterogeneous data at the current process stage. The probability of the following This indicates the state of multi-source heterogeneous data from adjacent process stages. The probability of the following, the first The multidimensional bin index corresponds to the first... There are several states. The probabilities of multi-source heterogeneous data in the current process stage under all states constitute the multi-source heterogeneous data distribution of the current process stage, and the probabilities of multi-source heterogeneous data in adjacent process stages under all states constitute the multi-source heterogeneous data distribution of adjacent process stages.

[0041] In one embodiment of this application, the process of converting multi-source heterogeneous data of the current process stage into probabilities of multi-source heterogeneous data of the current process stage under all states, and probabilities of multi-source heterogeneous data of adjacent process stages under all states, includes: cleaning and preprocessing the multi-source heterogeneous data of the current process stage to obtain preprocessed multi-source heterogeneous data of the current process stage; performing feature extraction (vibration data feature extraction, current data feature extraction, voltage data feature extraction, etc.) on the preprocessed multi-source heterogeneous data of the current process stage; cleaning and preprocessing the multi-source heterogeneous data of adjacent process stages to obtain preprocessed multi-source heterogeneous data of adjacent process stages; and performing feature extraction (vibration data feature extraction, current data feature extraction, voltage data feature extraction, etc.) on the preprocessed multi-source heterogeneous data of adjacent process stages. The vibration data feature range is determined based on the vibration data characteristics (obtained by merging vibration data characteristics extracted from multi-source heterogeneous data of the current process stage after preprocessing and vibration data characteristics extracted from multi-source heterogeneous data of adjacent process stages after preprocessing); the current data feature range is determined based on the current data characteristics (obtained by merging current data characteristics extracted from multi-source heterogeneous data of the current process stage after preprocessing and current data characteristics extracted from multi-source heterogeneous data of adjacent process stages after preprocessing); the voltage data feature range is determined based on the voltage data characteristics (obtained by merging voltage data characteristics extracted from multi-source heterogeneous data of the current process stage after preprocessing and voltage data characteristics extracted from multi-source heterogeneous data of adjacent process stages after preprocessing); the vibration data feature range is then divided into... The current data characteristic range is divided into several intervals (boxes). The voltage data characteristic range is divided into several intervals. The number of intervals is calculated by multiplying the number of intervals for the vibration data characteristic range, the number of intervals for the current data characteristic range, and the number of intervals for the voltage data characteristic range, to form... A multidimensional grid, each multidimensional grid representing a state. It represents a multidimensional interval composed of specific vibration data feature intervals, specific voltage data feature intervals, and specific current data feature intervals; it counts the number of data points in each multidimensional grid that the feature data extracted from the preprocessed multi-source heterogeneous data of the current process stage falls into, and uses the ratio of the number of data points in each multidimensional grid of the current process stage to the total data volume as the probability of the multi-source heterogeneous data of the current process stage in all states; it also counts the number of data points in each multidimensional grid that the feature data extracted from the preprocessed multi-source heterogeneous data of adjacent process stages falls into, and uses the ratio of the number of data points in each multidimensional grid of adjacent process stages to the total data volume as the probability of the multi-source heterogeneous data of adjacent process stages in all states.

[0042] In one embodiment of this application, if there are multiple categories of vibration data features, current data features, and voltage data features, the dimensions of the multidimensional grid will increase accordingly.

[0043] If the execution order is first execution followed by intermediate execution, and the remaining runtime is less than or equal to the duration of the transition state, then the current process stage is determined to be in a post-transition state. In one embodiment of this application, if the current process stage is in a post-transition state, then the adjacent process stage is the next process stage.

[0044] If the execution order is intermediate execution followed by final execution, and the elapsed execution time is less than the transition state duration, then the current process stage is determined to be in the pre-transition state. In one embodiment of this application, if the current process stage is in the pre-transition state, then the adjacent process stage is the previous process stage.

[0045] If the execution order is intermediate execution, the runtime already elapsed is greater than or equal to the transition state duration, and the remaining runtime is greater than the transition state duration, then the current process stage is determined to be in a stable state. In one embodiment of this application, the intermediate execution process stage has a pre-transition state duration and a post-transition state duration, and the total runtime of the current process stage is greater than twice the transition state duration.

[0046] If the execution order is last, the runtime already elapsed is greater than or equal to the transition state duration, and the remaining runtime is greater than the preset shutdown stage duration, then the current process stage is determined to be in a stable state. In one embodiment of this application, the preset shutdown stage duration is set according to the type of industrial production line and process stage. To reduce the occurrence of transition stages, the anomaly detection model set for the last executed process stage in the stable state is consistent with the anomaly detection model set for the last executed process stage in the shutdown stage, and the anomaly detection model set for the last executed process stage is set taking into account the shutdown stage.

[0047] In one embodiment of this application, the process of determining the anomaly score of the current process stage based on multi-source heterogeneous data of the current process stage, a first anomaly detection model set, and a first pre-configured weight combination includes:

[0048] Preprocessing is performed on the multi-source heterogeneous data in the current process stage to obtain preprocessed multi-source heterogeneous data. In one embodiment of this application, the preprocessing includes: data denoising, time alignment, sliding window segmentation, feature extraction, feature normalization, etc. Through the above processing, duplicate data and noise signals in the multi-source heterogeneous data can be removed, and data from different sources and with different sampling frequencies can be aligned to the same time dimension. Then, the data is divided into continuous data segments according to a preset window size, and the features of each data segment are extracted and normalized to eliminate the influence of features with different dimensions on the calculation of the anomaly detection model, which facilitates the unified calculation and inference of the subsequent anomaly detection model.

[0049] The preprocessed multi-source heterogeneous data is input into each anomaly detection model in the first anomaly detection model set to obtain a first initial anomaly score set. In one embodiment of this application, the first anomaly detection model set is determined according to the current process stage. The first anomaly detection model set only contains the numbers of the anomaly detection models participating in the operation, for example, {A, B}, where A represents an LSTM autoencoder and B represents an isolated forest. In this case, if the anomaly detection model not participating in the operation is a graph neural network model, the first pre-configured weight combination can be set to [0.7, 0.3, 0]. After the preprocessed multi-source heterogeneous data is input into each anomaly detection model in the first anomaly detection model set, the anomaly detection scores output by the participating anomaly detection models form the first initial anomaly score set. The number of scores in the first initial anomaly score set is consistent with the number of participating anomaly detection models. The LSTM autoencoder, isolated forest, and graph neural network models are all obtained through pre-training.

[0050] According to the first pre-configured weight combination, the scores in the first initial anomaly score set are weighted and fused to obtain the anomaly score. In one embodiment of this application, the formula for calculating the anomaly score is as follows:

[0051] Equation (3)

[0052] in, Indicates abnormal scores. Indicates the first Pre-configured weights for anomaly detection models Indicates the first The anomaly detection score output by each anomaly detection model is equal to 1, and the anomaly detection score output by an anomaly detection model that is not running is 0.

[0053] In one embodiment of this application, the process of determining the anomaly score of the current process stage based on multi-source heterogeneous data of the current process stage, a first anomaly detection model set, a first pre-configured weight combination, a second anomaly detection model set, and a second pre-configured weight combination includes:

[0054] Preprocessing is performed on the multi-source heterogeneous data in the current process stage to obtain preprocessed multi-source heterogeneous data. In one embodiment of this application, the preprocessing includes: data denoising, time alignment, sliding window segmentation, feature extraction, feature normalization, etc. Through the above processing, duplicate data and noise signals in the multi-source heterogeneous data can be removed, and data from different sources and with different sampling frequencies can be aligned to the same time dimension. Then, the data is divided into continuous data segments according to a preset window size, and the features of each data segment are extracted and normalized to eliminate the influence of features with different dimensions on the calculation of the anomaly detection model, which facilitates the unified calculation and inference of the subsequent anomaly detection model.

[0055] The anomaly detection models in the first anomaly detection model set and the anomaly detection model in the second anomaly detection model set are merged to obtain an updated anomaly detection model set. In one embodiment of this application, the anomaly detection models in the updated anomaly detection model set are unique; that is, if there is a duplicate anomaly detection model between the first and second anomaly detection model sets, the duplicate anomaly detection model is deleted to ensure the uniqueness of the anomaly detection model.

[0056] The preprocessed multi-source heterogeneous data is input into each anomaly detection model in the anomaly detection model update set to obtain a second initial anomaly score set. In one embodiment of this application, after the preprocessed multi-source heterogeneous data is input into each anomaly detection model in the anomaly detection model update set, the anomaly detection scores output by the participating anomaly detection models constitute the second initial anomaly score set. The number of scores in the second initial anomaly score set is consistent with the number of participating anomaly detection models.

[0057] According to the anomaly detection models in the anomaly detection model update set, the model weights in the first pre-configured weight combination and the model weights in the second pre-configured weight combination are interpolated to obtain the smooth transition weight of each anomaly detection model in the anomaly detection model update set. In one embodiment of this application, taking the current process stage as a post-transition state and the anomaly detection model as an LSTM autoencoder as an example, the calculation formula for the smooth transition weight of the LSTM autoencoder is as follows:

[0058] Equation (4)

[0059] in, This represents the smooth transition weights of the LSTM autoencoder. This represents the pre-configured weights of the LSTM autoencoder at the current process stage. Indicates the time point to switch. Indicates the duration of the transition state. Indicates a reference time point within the time range of the transition window. Inside, This represents the pre-configured weights of the LSTM autoencoder in the next process stage. The calculation method for the smooth transition weights of other anomaly detection models is the same as that for the LSTM autoencoder.

[0060] In one embodiment of this application, the duration of the transition state includes the length of multiple preset sliding windows. The process of determining the anomaly score of the current process stage is performed in units of preset sliding windows, based on multi-source heterogeneous data of the current process stage, a first set of anomaly detection models, a first set of pre-configured weights, a second set of anomaly detection models, and a second set of pre-configured weights.

[0061] In one embodiment of this application, if the feature data includes mean and variance, the process of determining a dynamic scoring threshold based on the feature data of multi-source heterogeneous data within a preset sliding window includes:

[0062] The mean of each dimension of data and the correlation between different dimensions of data are calculated within a preset number of sliding windows. In one embodiment of this application, the preset number of sliding windows can be 10 sliding windows or any other number of sliding windows. The preset number of sliding windows can be the 10 sliding windows preceding the current sliding window. The mean of each dimension of data forms a mean vector, and the expression for the mean vector is as follows:

[0063] Equation (5)

[0064] in, Represents the mean vector. This indicates the data length for each dimension. Indicates the first A multidimensional vector at a given time point, for example, a vector containing vibration data, voltage data, and current data.

[0065] In one embodiment of this application, the formula for calculating the correlation between data of different dimensions is as follows:

[0066] Equation (6)

[0067] in, The covariance matrix represents the correlation between data from different dimensions. This indicates the data length for each dimension. Indicates the first A multidimensional vector at each time point Represents the mean vector. This indicates transpose.

[0068] Based on the mean of data in each dimension and the correlation between data in different dimensions, the Mahalanobis distance of multi-source heterogeneous data within each sliding window is calculated. In one embodiment of this application, the sliding window for calculating the Mahalanobis distance belongs to a preset number of sliding windows. The formula for calculating the Mahalanobis distance of multi-source heterogeneous data within each sliding window is as follows:

[0069] Equation (7)

[0070] in, Indicates the first Mahalanobis distance of multi-source heterogeneous data within a sliding window Indicates the first A mean vector composed of the mean values ​​of each dimension of the multi-source heterogeneous data within a sliding window. Represents the mean vector. This represents the covariance matrix between data of different dimensions.

[0071] The mean of the Mahalanobis distance is calculated based on the Mahalanobis distance of multi-source heterogeneous data within each sliding window. In one embodiment of this application, the formula for calculating the mean of the Mahalanobis distance is as follows:

[0072] Equation (8)

[0073] in, This represents the mean of the distance to Maharanobis. This indicates the total number of sliding windows. Indicates the first Mahalanobis distance of multi-source heterogeneous data within a sliding window.

[0074] The standard deviation of the Mahalanobis distance is calculated based on the Mahalanobis distance and the mean of the Mahalanobis distances of the multi-source heterogeneous data within each sliding window. In one embodiment of this application, the formula for calculating the standard deviation of the Mahalanobis distance is as follows:

[0075] Equation (9)

[0076] in, This represents the standard deviation of the distance to Maharanobis. This represents the mean of the distance to Maharanobis. This indicates the total number of sliding windows. Indicates the first Mahalanobis distance of multi-source heterogeneous data within a sliding window.

[0077] A dynamic scoring threshold is calculated based on the mean and standard deviation of the Mahalanobis distance. In one embodiment of this application, the formula for calculating the dynamic scoring threshold is as follows:

[0078] Equation (10)

[0079] in, Indicates the dynamic scoring threshold. This represents the mean of the distance to Maharanobis. This represents the coefficient for adjusting sensitivity, which is set according to different process stages (e.g., during the welding process). (2.5) This represents the standard deviation of the Maharanobis distance.

[0080] In one embodiment of this application, determining the dynamic scoring threshold in the above manner can flexibly adjust the scoring criteria according to the distribution characteristics of production data at different process stages and time stages, avoiding the false alarms and missed alarms caused by changes in data distribution after process switching due to fixed thresholds, and effectively improving the accuracy and adaptability of abnormal detection results under different process stages.

[0081] In one embodiment of this application, after obtaining the cause of the abnormal state triggered in the current process stage, the process stage adaptive industrial production line anomaly detection method further includes:

[0082] The reasoning conditions and the reasons for triggering abnormal states in the current process stage are combined as rule cases to update the process diagnosis rules, which are then used for subsequent rule matching. In one embodiment of this application, only the newly added rule cases are stored in the rule base, eliminating the need to retrain all rules, thus ensuring the efficiency of rule updates. Simultaneously, it allows for the continuous accumulation of abnormal diagnosis experience at different process stages, gradually improving the accuracy of abnormal cause reasoning and adapting to changes brought about by process parameter adjustments and equipment aging during the long-term operation of industrial production lines.

[0083] Rule-based cases are used as incremental sample data, and the anomaly detection model is updated using this incremental sample data according to a preset update time. The updated anomaly detection model is then used to detect anomalies in subsequent process stages. In one embodiment of this application, the preset update time is set to midnight every day. The anomaly detection model is updated using incremental sample data from the previous day, improving its adaptability to changes in data characteristics under different operating conditions. This allows for continuous monitoring of production line operating status changes to optimize anomaly detection performance and avoids a decrease in detection accuracy due to prolonged periods without model updates. Models updated daily according to the preset update time include LSTM autoencoders. The LSTM autoencoder is updated through an incremental training process. During incremental training, the encoder in the LSTM autoencoder is frozen, and the parameters of the decoder are adjusted, taking approximately 4 minutes.

[0084] Figure 3 This is a schematic diagram of the architecture of an industrial production line anomaly detection system, as illustrated in an exemplary embodiment of this application. Figure 3The industrial production line anomaly detection system includes: a data acquisition and preprocessing module, a parallel inference module, a stage-aware routing module, a smooth transition module, a fusion judgment module, a multi-dimensional adaptive optimization module, and a feedback closed-loop module. The data acquisition and preprocessing module collects multi-source heterogeneous data and process stage labels (e.g., 1, 2, 3) from the industrial production line in real time, performs sliding window segmentation on the multi-source heterogeneous time-series data, and extracts and standardizes features from the multi-source heterogeneous data within each sliding window to obtain feature vectors. The stage-aware routing module determines the set of models that should participate in anomaly detection for the current process stage and the pre-configured weights of the participating models based on a preset "stage-model configuration" mapping table according to the current process stage label. The parallel inference module simultaneously inputs the feature vectors into the participating anomaly detection models, runs the participating anomaly detection models in parallel, and obtains the preliminary anomaly scores output by each model. The smooth transition module is used to perform multi-dimensional adaptive optimization on the production line. Within a preset time window before and after the segment switch, the anomaly detection models of the two stages are run simultaneously, and the pre-configured weights of the anomaly detection models are dynamically adjusted using time linear interpolation (or adaptive difference interpolation) to obtain smooth transition weights. The fusion judgment module is used to perform weighted fusion of the preliminary anomaly scores output by each model using the pre-configured weights of each model to obtain an anomaly score, or to perform weighted fusion of the preliminary anomaly scores output by the anomaly detection models of the two stages using smooth transition weights to obtain an anomaly score, and compare the anomaly score with the dynamic scoring threshold to output the anomaly judgment result. The multi-dimensional adaptive optimization module is used to dynamically adjust the scoring threshold using normal multi-source heterogeneous data within a preset sliding window of the current process stage, call the domain expert knowledge base to perform post-processing interpretation of the anomaly judgment result, and add the post-processing interpretation of the anomaly judgment result to the expert knowledge base, and use incremental sample data to learn and update the parameters of the LSTM autoencoder online.

[0085] Figure 4 This is an exemplary embodiment of the present application illustrating the relationship diagram of adaptive collaboration within a multidimensional adaptive optimization module. Figure 4 The multidimensional adaptive optimization module includes an online learning unit, a knowledge reasoning unit, and a dynamic thresholding unit. The online learning unit, after obtaining the post-processing interpretation of the anomaly determination results, uses rule cases as incremental sample data and employs this incremental sample data to learn and update the parameters of the LSTM autoencoder online. The knowledge reasoning unit is used to call upon a domain expert knowledge base to post-process and interpret the anomaly determination results when the current process stage is in an abnormal state. The dynamic thresholding unit dynamically adjusts the scoring threshold using normal multi-source heterogeneous data within a preset sliding window for the current process stage, thereby determining whether the current process stage is in an abnormal state based on the dynamic scoring threshold.

[0086] Figure 5This is a schematic diagram illustrating the dynamic change process of model combination and weights before and after stage switching, as shown in an exemplary embodiment of this application. Figure 5 Taking the transition from stage 1 to stage 2 as an example, during the two windows before and after the transition, anomaly detection is performed using an LSTM autoencoder and an isolated forest model in the stage before the transition, and anomaly detection is performed using an isolated forest and a graph neural network model in the stage after the transition. By performing linear interpolation calculation on the pre-configured weights of the anomaly detection model in stage 1 and the pre-configured weights of the anomaly detection model in stage 2 during the window period, the linear interpolation weights of the LSTM autoencoder, isolated forest, and graph neural network model show a smooth transition trend. Thus, after weighted fusion of the anomaly detection scores output by the LSTM autoencoder, isolated forest, and graph neural network model through linear interpolation weights, abrupt changes in the anomaly scores are avoided.

[0087] In one embodiment of this application, taking an automated welding production line for automotive parts as an example, the automated welding production line for automotive parts includes three process stages: loading (stage 1), welding (stage 2), and unloading (stage 3). The welding stage is a critical quality node, and abnormalities such as electrode tip wear, welding spatter, and insufficient cooling may occur. The hardware configuration required for this automated welding production line for automotive parts is as follows: an edge industrial control computer (equipped with an Intel Core i7-10700, 16GB RAM, and an NVIDIA T4 GPU), sensors (4 robot end effector vibration sensors (sampling frequency 1kHz), welding machine current / voltage sensors (sampling frequency 100Hz), PLC: outputting process stage labels (e.g., 1, 2, 3) via OPC UA protocol, with a latency of <1ms. The parallel inference module is deployed and runs on the edge industrial control computer.

[0088] The advantages of the adaptive industrial production line anomaly detection method described in this application for detecting anomalies in an automated welding production line for automotive parts, compared with existing technologies, are shown in Table 1:

[0089] Table 1

[0090]

[0091] As shown in Table 1, the method in this application has advantages over existing technologies, including high accuracy (F1 value above 0.95, an improvement of 6-7 percentage points), ultra-low false alarm rate during handover (false alarm rate during stage handover reduced to below 6%, and false alarm rate during handover reduced to below 2%, significantly reducing unnecessary downtime), and high real-time performance (avoiding full-model operation at all stages, with an average latency of 49ms, meeting millisecond-level requirements for production lines). Furthermore, the method in this application also features strong interpretability (diagnostic conclusions are output through post-processing using an expert knowledge base, allowing operators to understand routing decisions and the causes of anomalies) and resource conservation (CPU utilization is reduced by 42% compared to parallel operation of the full model, which is beneficial for edge deployment). Moreover, the performance of this application can be further enhanced by using an adaptive window (dynamically adjusting the smooth transition time based on the KL divergence of different process stages): further reducing the false alarm rate in extreme difference scenarios.

[0092] In one embodiment of this application, the half-window length (i.e., the duration of the transition state) can be adaptively adjusted by determining the half-window length based on the similarity between the multi-source heterogeneous data distribution of the current process stage and the multi-source heterogeneous data distribution of adjacent process stages. Experiments have verified that when the half-window length Δt = 2 sliding windows, the F1 score reaches a peak of 0.95 and the delay only increases by 10ms on the basis of the original delay time, which is a relatively ideal parameter. When Δt = 3 sliding windows, the delay increases too much and is not recommended as a parameter. When Δt = 1 sliding window, the false alarm rate of switching is still too high and is not recommended as a parameter. This proves that the anomaly detection effect is better when Δt = 2 sliding windows.

[0093] In one embodiment of this application, experimental verification shows that near the stage switching point, the F1 value of the method without a smooth transition scheme drops significantly, while the F1 value of the method with a smooth transition mechanism and the method with a smooth transition mechanism and the application of an adaptive window method remains stable, thus proving the effectiveness of the smooth transition mechanism.

[0094] The following describes an embodiment of the apparatus described in this application, which can be used to execute the process-stage adaptive industrial production line anomaly detection system described in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the process-stage adaptive industrial production line anomaly detection method described in the above embodiments of this application.

[0095] Figure 6 This is a block diagram illustrating an adaptive industrial production line anomaly detection system for process stages, as shown in an exemplary embodiment of this application.

[0096] like Figure 6 As shown, this exemplary process-stage adaptive industrial production line anomaly detection system 600 includes:

[0097] The data acquisition module 601 is used to acquire the current process stage of the production line, as well as the execution order, runtime, remaining runtime, and multi-source heterogeneous data of the current process stage;

[0098] The status determination module 602 is used to determine the first set of anomaly detection models and the first pre-configured weight combination of the first set of anomaly detection models according to the current process stage; and to determine the running status of the current process stage according to the execution order, the running time, and the remaining running time.

[0099] The steady state score determination module 603 is used to determine the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination if the current process stage is in a steady state.

[0100] The transition state score determination module 604 is used to determine the second anomaly detection model set and the second pre-configured weight combination of the second anomaly detection model set based on the adjacent process stages of the current process stage if the current process stage is in a transition state; and to determine the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, the first pre-configured weight combination, the second anomaly detection model set and the second pre-configured weight combination.

[0101] The cause reasoning module 605 is used to determine that the current process stage is in an abnormal state if the abnormal score is greater than the dynamic scoring threshold. The current process stage, the equipment source of the multi-source heterogeneous data of the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data of the current process stage are used as reasoning conditions. The module is matched with the pre-built process diagnosis rules or fault tree to obtain the cause of the abnormal state triggered by the current process stage.

[0102] In one embodiment of this application, the current process stage and its execution order are characterized by stage status codes in the status register of the programmable logic controller (PLC) on the industrial production line. For example, 1 represents the loading stage, 2 represents the welding stage, and 3 represents the unloading stage. Characterizing the current process stage using stage status codes in the PLC's status register eliminates the need for feature analysis calculations on sensor data, effectively reducing processing latency and keeping it within less than 1ms. The elapsed runtime is stored in the PLC after being timed by a timer. The remaining runtime is the difference between the total runtime and the elapsed runtime, and this remaining runtime is calculated and stored in the PLC. Multi-source heterogeneous data is collected by sensors and stored within the PLC. Vibration data is detected by vibration sensors, current data by current sensors, and voltage data by voltage sensors. Equipment requiring vibration data monitoring consists of key rotating machinery or transmission components in industrial production lines, including drive motors, gearboxes, fluid pumps, industrial fans, machine tool spindles, or welding positioners. Equipment requiring current data monitoring consists of power drive units in industrial production lines, including AC inverters, DC drivers, and servo drivers. Equipment requiring voltage data monitoring consists of core power supply units in industrial production lines, including incoming power distribution cabinet busbars and the output terminals of dedicated power transformers. All multi-source heterogeneous data can be directly read through the PLC's industrial Ethernet interface, ensuring real-time data retrieval.

[0103] In one embodiment of this application, the operating states include: a stable state, a transitional state, etc. The process of determining the first anomaly detection model set and the first pre-configured weight combination of the first anomaly detection model set according to the current process stage includes: upon obtaining the current process stage, searching for the first anomaly detection model set corresponding to the current process stage and the pre-configured weights of each anomaly detection model in the first anomaly detection model set through a pre-established "stage-model configuration" mapping table, and the pre-configured weights of each anomaly detection model in the first anomaly detection model set forming the first pre-configured weight combination. The operating states include: a stable state, a transitional state, etc. The above method of determining the first anomaly detection model set and the first pre-configured weight combination of the first anomaly detection model set according to the current process stage ensures that different process stages match the corresponding first anomaly detection model set and the first pre-configured weight combination of the first anomaly detection model set. This achieves dynamic switching of the detection model combination with the process stage, avoids insufficient performance of a single model throughout the entire process stage, and improves the flexibility and accuracy of anomaly detection.

[0104] In one embodiment of this application, the process of determining the anomaly score of the current process stage based on multi-source heterogeneous data of the current process stage, a first anomaly detection model set, and a first pre-configured weight combination includes: preprocessing the multi-source heterogeneous data of the current process stage to obtain preprocessed multi-source heterogeneous data; inputting the preprocessed multi-source heterogeneous data into each anomaly detection model in the first anomaly detection model set to obtain a first initial anomaly score set; and weighting and fusing the scores in the first initial anomaly score set according to the first pre-configured weight combination to obtain an anomaly score. This process, by comprehensively utilizing multi-source heterogeneous data to obtain a first initial anomaly score set output by multiple models, and weighting and fusing the scores in the first initial anomaly score set output by multiple models to obtain an anomaly score, fully leverages the detection advantages of different anomaly detection models for different anomaly types, further improving the accuracy of anomaly detection.

[0105] In one embodiment of this application, the transition state includes a pre-transition state and a post-filtering state. If the current process stage is in the pre-transition state, a second anomaly detection model set and a second pre-configured weight combination of the second anomaly detection model set are determined based on the previous process stage of the current process stage. Based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, the first pre-configured weight combination, the second anomaly detection model set and the second pre-configured weight combination, the anomaly score of the current process stage is determined, thus completing the smooth transition between the previous process stage and the current process stage. If the current process stage is in a post-transition state, a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models are determined based on the next process stage. Then, based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, the anomaly score of the current process stage is determined. After determining the anomaly score of the current process stage, the anomaly score of the next process stage is determined based on the multi-source heterogeneous data of the next process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models, and the second pre-configured weight combination, thus completing the smooth transition between the current and next process stages. The process of determining the anomaly score of the next process stage based on the multi-source heterogeneous data of the next process stage is the same as the process of determining the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage.

[0106] In one embodiment of this application, the dynamic scoring threshold is determined based on the feature data of multi-source heterogeneous data within a preset sliding window. The feature data of the vibration data includes: time-domain statistical features, frequency-domain features, and time-frequency-domain features. The time-domain statistical features include: root mean square, peak value, skewness, and kurtosis. The frequency-domain features are extracted through Fourier transform and include: energy proportion of each frequency band, centroid frequency, and frequency variance. The time-frequency-domain features are obtained through wavelet decomposition and include: energy statistical values ​​of each wavelet component. The feature data of the current data includes: average current, current fluctuation variance, and harmonic component amplitude proportion. The feature data of the voltage data includes: effective voltage value, voltage fluctuation rate, and voltage deviation rate. Pre-built process diagnostic rules or fault trees are stored in an expert knowledge base. Rule matching with the pre-built process diagnostic rules or fault trees is implemented through an inference engine. The inference engine can be selected according to the actual situation, such as a large language inference model or a fault tree modeling and simulation platform for industrial control systems.

[0107] It should be noted that the process-stage adaptive industrial production line anomaly detection system and the process-stage adaptive industrial production line anomaly detection method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the process-stage adaptive industrial production line anomaly detection system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0108] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, enable the electronic device to implement the process-stage adaptive industrial production line anomaly detection method provided in the various embodiments above.

[0109] Figure 7 A schematic diagram of a computer system suitable for implementing the embodiments of this application is shown. It should be noted that... Figure 7 The computer system 700 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0110] like Figure 7As shown, the computer system 700 includes a Central Processing Unit (CPU) 701, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on programs stored in Read-Only Memory (ROM) 702 or programs loaded from storage portion 708 into Random Access Memory (RAM) 703. The RAM 703 also stores various programs and data required for system operation. The CPU 701, ROM 702, and RAM 703 are interconnected via a bus 704. An Input / Output (I / O) interface 705 is also connected to the bus 704.

[0111] The following components are connected to I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to I / O interface 705 as needed. A removable medium 711, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 710 as needed so that computer programs read from it can be installed into storage section 708 as needed.

[0112] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 709, and / or installed from removable medium 711. When the computer program is executed by central processing unit (CPU) 701, it performs various functions defined in the system of this application.

[0113] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium, a computer-readable medium, or any combination of the above. For example, a computer-readable medium can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than a computer-readable medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. A computer program contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0114] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0115] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0116] Another aspect of this application provides a computer-readable medium storing a computer program that, when executed by a computer's processor, causes the computer to perform an industrial production line anomaly detection method adapted to the aforementioned process stages. This computer-readable medium may be included in the electronic device described in the above embodiments, or it may exist independently and not incorporated into that electronic device.

[0117] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable medium. A processor of a computer device reads the computer instructions from the computer-readable medium and executes the computer instructions, causing the computer device to perform the process-stage adaptive industrial production line anomaly detection method provided in the various embodiments above.

[0118] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. An adaptive industrial production line anomaly detection method for process stages, characterized in that, The method includes: The current process stage of the production line is obtained, along with the execution sequence, elapsed runtime, remaining runtime, and multi-source heterogeneous data of the current process stage; the multi-source heterogeneous data of the current process stage includes: vibration data, current data, and voltage data; Based on the current process stage, a first set of anomaly detection models and a first pre-configured weight combination of the first set of anomaly detection models are determined; based on the execution order, the runtime elapsed, and the remaining runtime, the operating state of the current process stage is determined; the operating state includes: a stable state and a transitional state. If the current process stage is in a stable state, then the anomaly score of the current process stage is determined based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination. If the current process stage is in a transitional state, then based on the adjacent process stages of the current process stage, a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models are determined; and based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models and the second pre-configured weight combination, the anomaly score of the current process stage is determined. If the abnormal score is greater than the dynamic scoring threshold, the current process stage is determined to be in an abnormal state. The current process stage, the equipment source of the multi-source heterogeneous data of the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data of the current process stage are used as inference conditions and matched with pre-built process diagnosis rules or fault trees to obtain the reason for the abnormal state triggered by the current process stage. The dynamic scoring threshold is determined based on the feature data of the multi-source heterogeneous data within a preset number of sliding windows.

2. The adaptive industrial production line anomaly detection method according to claim 1, characterized in that, If the transition state includes a pre-transition state and a post-filtering state, then the process of determining the operating state of the current process stage based on the execution order, the elapsed runtime, and the remaining runtime includes: If the execution order is the first execution, the runtime is greater than or equal to the preset cold start state duration, and the remaining runtime is greater than the transition state duration, then the current process stage is determined to be in the stable state; the transition state duration is determined based on the similarity between the multi-source heterogeneous data distribution of the current process stage and the multi-source heterogeneous data distribution of adjacent process stages. If the execution order is first execution and middle execution, and the remaining runtime is less than or equal to the duration occupied by the transition state, then the current process stage is determined to be in the post-transition state. If the execution order is middle execution and end execution, and the running time is less than the duration occupied by the transition state, then the current process stage is determined to be in the pre-transition state. If the execution order is intermediate execution, the runtime of the current process is greater than or equal to the duration of the transition state, and the remaining runtime is greater than the duration of the transition state, then the current process stage is determined to be in the stable state. If the execution order is last, the runtime of the current process is greater than or equal to the duration of the transition state, and the remaining runtime is greater than the duration of the preset shutdown phase, then the current process stage is determined to be in the stable state.

3. The adaptive industrial production line anomaly detection method according to claim 2, characterized in that, If the similarity between the multi-source heterogeneous data distribution of the current process stage and the multi-source heterogeneous data distribution of the adjacent process stages is characterized by KL divergence, then the formula for calculating the duration of the transition state is as follows: include: ， in, Indicates the duration of the transition state. Indicates the basic usage time. Indicates the stretching factor. Denotes KL divergence, This indicates the distribution of multi-source heterogeneous data at the current process stage. This represents the distribution of multi-source heterogeneous data in adjacent process stages. Indicates the shortest transition half-window. This indicates the longest transition half-window.

4. The adaptive industrial production line anomaly detection method according to any one of claims 1-3, characterized in that, The process of determining the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination includes: The multi-source heterogeneous data of the current process stage are preprocessed to obtain preprocessed multi-source heterogeneous data; the preprocessing includes: data denoising, time alignment, sliding window separation, feature extraction, and feature normalization. The preprocessed multi-source heterogeneous data is input into each anomaly detection model in the first anomaly detection model set to obtain the first initial anomaly score set. According to the first pre-configured weight combination, the scores in the first initial abnormal score set are weighted and fused to obtain the abnormal score.

5. The adaptive industrial production line anomaly detection method according to any one of claims 1-3, characterized in that, The process of determining the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, the first pre-configured weight combination, the second anomaly detection model set, and the second pre-configured weight combination includes: The multi-source heterogeneous data of the current process stage are preprocessed to obtain preprocessed multi-source heterogeneous data; the preprocessing includes: data denoising, time alignment, sliding window separation, feature extraction, and feature normalization. The anomaly detection models in the first anomaly detection model set and the anomaly detection models in the second anomaly detection model set are merged to obtain an anomaly detection model update set; the anomaly detection models in the anomaly detection model update set are unique. The preprocessed multi-source heterogeneous data is input into each anomaly detection model in the anomaly detection model update set to obtain a second initial anomaly score set. According to the anomaly detection model in the anomaly detection model update set, the model weights in the first pre-configured weight combination and the model weights in the second pre-configured weight combination are interpolated to obtain the smooth transition weight of each anomaly detection model in the anomaly detection model update set. Based on the smooth transition weight of each anomaly detection model in the anomaly detection model update set, the scores in the second initial anomaly score set are weighted and fused to obtain the anomaly score.

6. The adaptive industrial production line anomaly detection method according to any one of claims 1-3, characterized in that, If the feature data includes mean and variance, then the process of determining the dynamic scoring threshold based on the feature data of multi-source heterogeneous data within a preset sliding window includes: Calculate the mean of data in each dimension and the correlation between data in different dimensions within the preset number of sliding windows; Based on the mean of data in each dimension and the correlation between data in different dimensions, calculate the Mahalanobis distance of multi-source heterogeneous data within each sliding window. The mean of the Mahalanobis distance is calculated based on the Mahalanobis distance of multi-source heterogeneous data within each sliding window; The standard deviation of the Mahalanobis distance is calculated based on the Mahalanobis distance of the multi-source heterogeneous data within each sliding window and the mean of the Mahalanobis distance. The dynamic scoring threshold is calculated based on the mean and standard deviation of the Mahalanobis distance.

7. The adaptive industrial production line anomaly detection method according to any one of claims 1-3, characterized in that, After obtaining the cause of the abnormal state triggered in the current process stage, the method further includes: The reasoning conditions and the reasons for the abnormal state triggered in the current process stage are combined as rule cases to update the process diagnosis rules, so as to perform subsequent rule matching based on the updated process diagnosis rules. The rule cases are used as incremental sample data, and the anomaly detection model is updated using the incremental sample data according to a preset update time, so as to perform anomaly detection in subsequent process stages using the updated anomaly detection model.

8. An adaptive industrial production line anomaly detection system for process stages, characterized in that, include: The data acquisition module is used to acquire the current process stage of the production line, as well as the execution order, runtime, remaining runtime, and multi-source heterogeneous data of the current process stage. The multi-source heterogeneous data in the current process stage includes: vibration data, current data, and voltage data; The state determination module is used to determine a first set of anomaly detection models and a first pre-configured weight combination of the first set of anomaly detection models based on the current process stage; and to determine the running state of the current process stage based on the execution order, the elapsed runtime, and the remaining runtime; the running state includes: a stable state and a transitional state. The steady state score determination module is used to determine the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first anomaly detection model set, and the first pre-configured weight combination if the current process stage is in a steady state. The transition state score determination module is used to determine, if the current process stage is in a transition state, a second set of anomaly detection models and a second pre-configured weight combination of the second set of anomaly detection models based on the adjacent process stages of the current process stage; and to determine the anomaly score of the current process stage based on the multi-source heterogeneous data of the current process stage, the first set of anomaly detection models, the first pre-configured weight combination, the second set of anomaly detection models and the second pre-configured weight combination. The cause reasoning module is used to determine that the current process stage is in an abnormal state if the abnormal score is greater than the dynamic scoring threshold. The current process stage, the equipment source of the multi-source heterogeneous data of the current process stage, and the feature vector composed of feature data extracted from the multi-source heterogeneous data of the current process stage are used as reasoning conditions to perform rule matching with pre-built process diagnosis rules or fault trees to obtain the cause of the abnormal state triggered by the current process stage. The dynamic scoring threshold is determined based on the feature data of the multi-source heterogeneous data within a preset number of sliding windows.

9. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement an industrial production line anomaly detection method that is process-stage adaptive as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-readable instructions, which, when executed by the computer's processor, cause the computer to perform the process-adaptive industrial production line anomaly detection method as described in claims 1 to 7.

Images