Alarm method, program product and storage medium for industrial equipment

Abstract

The present application relates to the industrial field, and provides an alarm method, program product, and storage medium for industrial equipment. Among them, the alarm method for industrial equipment includes: obtaining first original data, the first original data includes quality data of products produced by industrial equipment; obtaining first data to be analyzed based on the first original data, the first data to be analyzed includes quality feature data, and the quality feature data is obtained by feature extraction of quality data; matching the first data to be analyzed with the alarm rules in the rule base, if the match is successful, outputting the first alarm information based on the matched alarm rules, the first alarm information includes fault information and / or maintenance information of the industrial equipment. This method is conducive to improving the accuracy of fault alarms and / or maintenance alarms for industrial equipment, and ensuring the quality of the products produced.

Description

Technical Field

[0001] The present invention relates to the field of industrial technology, and in particular to an alarm method, a program product and a storage medium for industrial equipment. Background Art

[0002] At present, the degree of automation in industrial production is getting higher and higher, and a large number of equipment will be involved in the production line of products. For example, in the process of manufacturing photovoltaic modules, equipment such as dicing machines, stringing machines, stack welding machines, and laminating machines are required. Therefore, once a device fails during the production process, it will seriously affect the quality of the products produced. At present, product quality can be detected through computer vision technology, but there is still a lack of high-accuracy diagnosis or early warning methods for equipment failures. Summary of the invention

[0003] The purpose of the embodiments of the present application is to provide an alarm method, program product and storage medium for industrial equipment to improve the above technical problems.

[0004] To achieve the above objectives, this application provides the following technical solutions:

[0005] In a first aspect, an embodiment of the present application provides an alarm method for industrial equipment, comprising: acquiring first original data, the first original data including quality data of products produced by the industrial equipment; obtaining first data to be analyzed based on the first original data; wherein the first data to be analyzed includes quality feature data, and the quality feature data is obtained by feature extraction of the quality data; matching the first data to be analyzed with alarm rules in a rule base, and if the match is successful, outputting first alarm information based on the matched alarm rules; wherein the first alarm information includes fault information and / or maintenance information of the industrial equipment.

[0006] In an implementation of the first aspect, the method further includes: receiving first user feedback on the first alarm information, and iteratively optimizing the alarm rules in the rule base according to the first user feedback; wherein the first user feedback includes feedback on the authenticity of the first alarm information.

[0007] In an implementation of the first aspect, the method further includes: during the operation of the industrial equipment, obtaining second original data, the second original data including process data of the industrial equipment and / or operation status data of the industrial equipment; obtaining second data to be analyzed based on the second original data; performing anomaly detection on the second data to be analyzed using a first anomaly detection model to obtain an anomaly detection result; wherein the first anomaly detection model is a model obtained by unsupervised training; if the anomaly detection result indicates that the second data to be analyzed is abnormal, outputting second alarm information based on the anomaly detection result; wherein the second alarm information includes at least one of the following: data anomaly information, fault information of the industrial equipment associated with the data anomaly, and maintenance information of the industrial equipment associated with the data anomaly.

[0008] In an implementation of the first aspect, the method also includes: receiving second user feedback for the second alarm information, and constructing a labeled training sample based on the second user feedback and the second data to be analyzed; wherein the second user feedback includes feedback on the authenticity of the second alarm information; when the number of the collected training samples meets the training requirements, using the training samples to perform supervised training to obtain a second anomaly detection model; obtaining third original data, which is subsequent data of the second original data; obtaining third data to be analyzed based on the third original data; using the second anomaly detection model to perform anomaly detection on the third data to be analyzed to obtain anomaly detection results; or, combining the second anomaly detection model and the first anomaly detection model to perform anomaly detection on the third data to be analyzed to obtain anomaly detection results.

[0009] In a second aspect, an embodiment of the present application provides an alarm device for industrial equipment, comprising: a data acquisition module, used to acquire first original data, wherein the first original data includes quality data of products produced by the industrial equipment; a data processing module, used to obtain first data to be analyzed based on the first original data; wherein the first data to be analyzed includes quality feature data, and the quality feature data is obtained by feature extraction of the quality data; an alarm module, used to match the first data to be analyzed with the alarm rules in a rule base, and if the match is successful, output a first alarm information based on the matched alarm rules; wherein the first alarm information includes fault information and / or maintenance information of the industrial equipment.

[0010] In a third aspect, an embodiment of the present application provides a computer program product, including computer program instructions, which, when read and executed by a processor, execute the method provided by the first aspect or any one of the implementations of the first aspect.

[0011] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are read and executed by a processor, the method provided by the first aspect or any one of the implementations of the first aspect is executed.

[0012] In a fifth aspect, an embodiment of the present application provides an electronic device, comprising: a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are read and run by the processor, the method provided by the first aspect or any one of the implementations of the first aspect is executed.

[0013] Since the product is produced by industrial equipment, once the industrial equipment has a fault or needs maintenance, it will be reflected in the quality of the product. Therefore, in the above technical solution, the fault alarm and / or maintenance alarm of the equipment is performed based on the first original data containing the quality data of the product, which can improve the accuracy of the alarm, and is conducive to timely discovery of equipment failure or timely maintenance of the equipment, thereby ensuring the quality of the produced products. In addition, the scheme is not directly based on the first original data for alarm, but based on the first data to be analyzed further obtained from the first original data for alarm. Since the first data to be analyzed at least contains the quality feature data obtained by feature extraction from the quality data, that is, the features associated with the equipment failure or maintenance requirements in the first original data are more deeply excavated, it is also conducive to improving the accuracy of the alarm. In addition, based on this scheme, not only can a fault alarm be performed when a fault occurs, but also predictive maintenance can be performed when no fault occurs, which is conducive to reducing the probability of fault occurrence. In addition, the same processing logic is used for fault alarm and maintenance alarm in this scheme. For users, whether it is a fault alarm or a maintenance alarm, they only need to configure the corresponding alarm rules in the rule base, which is more convenient to handle.

[0014] In one implementation of the above technical solution, the alarm rules in the rule base can be iteratively optimized based on user feedback, which is conducive to integrating more manual experience into the rules and improving the accuracy of alarms.

[0015] In one implementation of the above technical solution, it is also supported to perform anomaly detection based on the acquired second original data during the operation of the equipment, so as to timely discover anomalies in the data and issue an alarm. Since the data anomaly may be caused by equipment failure or may be a sign that the equipment needs maintenance, this implementation method is conducive to early detection of equipment failure or maintenance of the equipment (rather than waiting until the product is produced). In addition, anomaly detection uses the first anomaly detection model obtained by unsupervised training, that is, there is no need to collect labeled training samples in advance, which is conducive to performing such anomaly detection in the early stage of production line operation and in the absence of data accumulation, and can also detect equipment failure or maintain the equipment early.

[0016] In one implementation of the above technical solution, a sufficient number of training samples (including labels) can be constructed based on user feedback to perform supervised training to obtain a second anomaly detection model. Since the second anomaly detection model absorbs more manual experience, it is likely to be better than the first anomaly detection model obtained based on unsupervised training in terms of performance. Therefore, when performing anomaly detection on the subsequent third original data, the first anomaly detection model can be replaced by the second anomaly detection model to improve the accuracy of anomaly detection. Of course, the first anomaly detection model can also be retained, and the two anomaly detection models can be combined for anomaly detection to obtain more reliable anomaly detection results. For example, for some accidental anomalies, if they are not included in the labeled training samples, the second anomaly detection model may perform poorly, while the first anomaly detection model may perform better. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A flowchart of an alarm method for industrial equipment provided in an embodiment of the present application;

[0019] Figure 2 A flowchart of another alarm method for industrial equipment provided in an embodiment of the present application;

[0020] Figure 3 A module diagram of an alarm device for industrial equipment provided in an embodiment of the present application;

[0021] Figure 4 A structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0023] The terms "comprises", "comprising" or any other variation thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. The terms "first", "second", "third", etc. are used only to distinguish descriptions and are not to be understood as indicating or implying relative importance.

[0024] Figure 1 A flowchart of an alarm method for industrial equipment provided in an embodiment of the present application. The method may be, but is not limited to, Figure 4 The electronic device shown is executed, and the structure of the electronic device can be referred to later. Figure 4 Reference Figure 1 , the alarm methods for industrial equipment include:

[0025] Step S110: Acquire first original data.

[0026] The first original data is the basic data used for equipment alarm analysis. The so-called "original" is relative to the first data to be analyzed in step S120. The first data to be analyzed may be processed on the basis of the first original data, and therefore is no longer "original".

[0027] The first original data includes at least the quality data of the products produced by the industrial equipment. The industrial equipment here refers to the equipment targeted by the alarm method, and the product refers to the object processed by the industrial equipment. It should be noted that the product here is not necessarily the final finished product to be manufactured, but can also be a semi-finished product, because there may be many industrial equipment distributed on the industrial assembly line, except for the terminal equipment, the rest of the equipment is only processing semi-finished products. For example, for the production of photovoltaic modules, the industrial equipment here can be equipment such as dicing machines, stringing machines, stacking machines, laminating machines, etc., and the product can be a photovoltaic module (including semi-finished photovoltaic modules).

[0028] Quality data characterizes the quality of a product. For example, quality data may include one or more data such as product defect data and product performance data. Quality data can be obtained by analyzing products using algorithms or measuring products using special devices after industrial equipment produces products. Depending on the implementation method, the quality data of a product can be obtained through analysis after each product is produced by industrial equipment, or the quality data of a batch of products can be obtained through analysis after the industrial equipment produces a batch of products, or these two methods can be combined to obtain quality data at different levels.

[0029] In addition to quality data, the first raw data may also include other data, such as data related to industrial equipment, and specific examples will be given later. In addition, the acquisition method and timing of other data may be different from that of quality data, for example, it may be collected during the process of industrial equipment producing products.

[0030] For example, on an industrial assembly line, quality inspection equipment can be set up at different workstations to perform quality inspections on products produced by the industrial equipment in different ways to obtain product quality data, which is then sent to a remote server, where the alarm method provided by the embodiment of the present application is executed. Of course, the server can also receive other data from the industrial equipment itself or external sensors.

[0031] Step S120: obtaining first data to be analyzed according to the first original data.

[0032] The first raw data can be processed to obtain the first data to be analyzed. The processing here can include at least one of the following two categories: the first category is preprocessing, the main purpose of which is to improve data quality; the second category is further processing, the main purpose of which is to convert the first raw data into a form that is more suitable for alarm analysis. In different implementations, all the first raw data can be processed to obtain the corresponding first data to be analyzed; only part of the first raw data can be processed to obtain the corresponding first data to be analyzed, and the remaining first raw data can be directly used as the first data to be analyzed; all the first raw data can also be directly used as the first data to be analyzed (this is equivalent to not processing the first raw data, which is a special case). In addition, even if some of the first raw data have been processed, the first raw data before and after the processing can be used as the first data to be analyzed.

[0033] For preprocessing of the first original data, the processing method may include, but is not limited to, at least one of the following:

[0034] (1) Missing value processing

[0035] Missing values ​​refer to blank data caused by abnormal reasons. For example, missing values ​​in a short period of time can be filled, such as filling with fixed values, using interpolation to fill, using data near the missing values ​​for prediction and filling, etc.; for missing values ​​in a long period of time, this part of the data can be directly discarded to avoid low-quality data affecting the accuracy of alarm analysis.

[0036] "Short time" and "long time" can be defined by the number of consecutive missing values. For example, if there are more than 10 consecutive missing values, it is considered to be a long-term missing situation, and if there are no more than 10 consecutive missing values, it is considered to be a short-term missing situation, and so on.

[0037] (2) Noise processing

[0038] There are different processing methods for different noises. For example, for some short-term noise, moving average can be used to smooth and remove it; for some high-frequency noise, low-pass filter can be used to remove it.

[0039] (3) Standardization

[0040] Standardization refers to converting diverse data into a unified form to facilitate subsequent operations. The "unification" here can refer to the unification of data dimensions, the unification of data value ranges, etc. For example, data can be standardized using the Z-score algorithm.

[0041] Further processing of the first raw data can be achieved by feature extraction, for example, feature data of the first raw data in one or more dimensions can be extracted, and these feature data can have a better correspondence with equipment failure or equipment maintenance requirements, thereby facilitating subsequent alarm analysis. Among them, equipment failure refers to abnormal working state of the equipment or even failure to work at all, while equipment maintenance requirements refer to the equipment is still working normally, but has shown some bad signs, so maintenance is required, otherwise it may cause failure.

[0042] In particular, since the first original data includes quality data, feature extraction can be performed on the quality data, and the obtained feature data is called quality feature data. Feature extraction can adopt, for example, statistical methods or model-based algorithms.

[0043] Step S130: Match the first data to be analyzed with the alarm rules in the rule base. If the match is successful, output the first alarm information based on the matched alarm rules.

[0044] Before executing step S130, the rule base has been established in advance. The rule base is a collection of a large number of alarm rules. These alarm rules can be constructed based on the correlation between the historical data to be analyzed and the real equipment failure or the real equipment maintenance requirements. The construction method can be, for example, automatic construction by algorithm, manual analysis construction, or a combination of these two methods. The historical data to be analyzed and the first data to be analyzed have similar meanings and are not explained in detail, but they are obtained before the first data to be analyzed. The historical data to be analyzed may come from the industrial equipment to be analyzed currently (or the products it produces), or may come from the same model of equipment as the industrial equipment to be analyzed currently (or the products it produces).

[0045] The alarm rules may include rules of the time dimension and / or rules of the non-time dimension. The rules of the time dimension are for the first data to be analyzed within a time window. For example, if the frequency of abnormalities in the data to be analyzed within a time window exceeds a certain threshold, it is determined that the equipment has a certain fault. The time window here can refer to a specific duration, such as 5 minutes, 10 minutes, or a certain number of products, such as every 10 products, every 100 products, or a certain batch of products, such as every batch of products, every two batches of products, and so on. In short, the time window does not necessarily correspond directly to the duration, and the number of products can also be regarded as another way of expressing the duration. When the time window is mentioned later, it can be understood in a similar way and will not be repeated. The rules of the non-time dimension are for a single first data to be analyzed. For example, if a first data to be analyzed is greater than a certain threshold, it is determined that the equipment needs some maintenance. Some examples of specific rules will be given later.

[0046] Whether it is a rule in the time dimension or a rule in the non-time dimension, there may be some parameters (such as thresholds), which can be initially given by experience or model prediction. And optionally, these parameters can also be continuously optimized as the alarm method is executed. Specific examples will be given later.

[0047] The alarm rules in the rule base may be fixed or variable. In addition to the parameter updates mentioned above, rules may also be added, modified, or deleted.

[0048] Matching the first data to be analyzed with the alarm rules in the rule base means verifying whether the first data to be analyzed conforms to the alarm rules one by one. Of course, different rule bases can be set for different industrial equipment, or only the alarm rules related to the equipment in the rule base can be matched; for different first data to be analyzed, only the alarm rules related to the first data to be analyzed in the rule base can be matched.

[0049] The alarm rules can record the fault information of industrial equipment, such as the cause of the fault, such as which component in the equipment caused the fault, or which improper operation or setting of the equipment caused the fault, and the solution to the fault, etc. Therefore, once such an alarm rule is matched, the first alarm information containing the fault information can be output. In addition to the cause of the fault and the fault solution, the first alarm information can also contain a fault prompt, that is, prompting the user that the current equipment has a fault.

[0050] The alarm rules may also record maintenance information of industrial equipment, such as maintenance suggestions, such as how to perform equipment maintenance, when to perform equipment maintenance, and the consequences of not performing maintenance, such as what kind of failures may be caused, etc. Therefore, once such an alarm rule is matched, the first alarm information containing maintenance information may be output. In addition to maintenance suggestions and the consequences of not performing maintenance, the first alarm information may also contain maintenance prompts, that is, prompting the user that the current equipment needs maintenance.

[0051] The alarm rules that record fault information are called fault rules, and the alarm rules that record maintenance information are called maintenance rules. The rule base can implement both fault rules and maintenance rules, or only one of the two types of rules can be implemented according to needs.

[0052] There are many ways to output the first alarm information, for example, it can be output to a system interface (for example, a web page or client interface) for display, or it can be output through voice broadcast, or it can be sent to the user through text messages or emails, or it can be output through light alarms, or it can also combine several different output methods at the same time.

[0053] If the first data to be analyzed fails to match any alarm rule, no processing is performed.

[0054] Brief summary Figure 1 In the method, since the product is produced by industrial equipment, once the industrial equipment fails or needs maintenance, it will cause a certain degree of decline in equipment performance, which will be reflected in the quality of the product. Therefore, in the above method, based on the first original data containing the quality data of the product to issue a fault alarm and / or a maintenance alarm for the equipment, the accuracy of the alarm can be improved, which is conducive to timely detection of equipment failures or timely maintenance of the equipment, thereby ensuring the quality of the products produced. In contrast, in the prior art, the quality data of the product is usually only used for product quality assessment, and is not associated with equipment status assessment.

[0055] In addition, the method does not issue an alarm directly based on the first original data, but issues an alarm based on the first data to be analyzed further obtained from the first original data. Since the first data to be analyzed at least includes quality feature data obtained by feature extraction from quality data, that is, it more deeply explores the features associated with equipment failure or maintenance requirements in the first original data, it is also beneficial to improve the accuracy of the alarm.

[0056] In addition, based on this method, not only can fault alarms be issued when a fault occurs, but predictive maintenance can also be performed when no fault occurs, which helps reduce the probability of faults occurring.

[0057] In addition, the method adopts the same processing logic for fault alarms and maintenance alarms. For users, no matter it is a fault alarm or a maintenance alarm, they only need to configure the corresponding alarm rules in the rule base, which is convenient to handle.

[0058] Based on the above embodiment, optionally, the quality data may further include at least one of the following two data:

[0059] (1) The results of defect detection on product images using computer vision algorithms.

[0060] The computer vision algorithm here can be a traditional image processing algorithm (for example, edge detection, morphological processing algorithm) or an artificial intelligence algorithm (for example, an algorithm based on a neural network model). These algorithms take product images as input and output defect detection results for the product images. The detection results can include the type of defect and the location of the defect in the image (for example, a detection box).

[0061] Products produced by normal equipment are usually free of defects. Therefore, if a product has defects, it can be inferred that there is something wrong with the equipment, such as a failure or urgent maintenance. Therefore, the data in category (1) is of great significance for equipment alarms.

[0062] (2) Numerical data obtained by measurement that characterizes the performance or specifications of a product.

[0063] Product performance, such as electrical performance, mechanical performance, etc. Electrical performance includes one or more values ​​such as voltage, current, power consumption, etc. Physical performance includes one or more values ​​such as strength, elasticity, hardness, etc. Product specifications include one or more values ​​such as product thickness, length, width, angle, etc. Methods for measuring these numerical data include: using professional measuring equipment for measurement; based on product images, using computer vision algorithms for measurement, such as first measuring the pixel distance between two points in the image, and then converting it into the distance between two points on the actual product.

[0064] The above numerical data represent the actual performance or specifications of the product, and the product also has preset requirements for performance or specifications when it is manufactured. By comparing the two, it is possible to make a judgment on failure or maintenance. Therefore, the second type of data is also important for equipment alarms.

[0065] In the above implementation, the quality data may come from the visual inspection results of the product, or may come from the numerical measurement results of the product. Rich types of quality data are conducive to improving the accuracy or comprehensiveness of the alarm.

[0066] Based on the above embodiments, optionally, the process of extracting quality feature data based on product quality data in step 120 may further include: extracting features from the quality data from at least one dimension of form, time and space to obtain quality feature data of the corresponding dimension.

[0067] Among them, the quality characteristic data of the morphological dimension mainly refers to the subdivided categories of defects represented by the quality data, the quality characteristic data of the time dimension mainly refers to the characteristics of the quality data over time, and the quality characteristic data of the spatial dimension mainly refers to the spatial location characteristics of the defects represented by the quality data.

[0068] In the above implementation, feature extraction of quality data from different dimensions is supported, and multi-dimensional quality feature data is conducive to improving the accuracy or comprehensiveness of alarms.

[0069] The following is an introduction to possible feature extraction methods for the above three dimensions. It should be understood that the feature extraction methods are not limited to the methods mentioned below:

[0070] A. Morphological Dimension

[0071] If the quality data includes the results of defect detection on product images using computer vision algorithms, the defect information in the detection results can be used to perform secondary classification of the product according to the product image, and the obtained classification results can be used as the quality feature data of the morphological dimension. The defect information can include the category and location of the defect, and the location of the defect can be, for example, a detection box.

[0072] For example, if the defect category of the defect detection result is that there is a hidden crack in the battery cell (a type of defect category), the defect area can be cropped out from the product image, and based on the cropped local image, the traditional image processing algorithm or artificial intelligence algorithm is used to further classify the hidden crack defects, such as hidden cracks in the main grid of the battery cell, hidden cracks on the upper edge, hidden cracks on the lower edge, X-shaped hidden cracks, V-shaped hidden cracks, etc., and these more specific classification results are used as quality feature data of the morphological dimension. The corresponding alarm rules are, for example: if the defect category is hidden cracks in the main grid, then there is a fault a in the industrial equipment. When the computer vision algorithm performs visual defect detection on the product, its detection result may include the defect category, that is, a defect classification is performed, but the classification result may be relatively rough and fails to correspond well to the cause of the equipment failure or the equipment maintenance requirement. Therefore, in the above implementation method, the product defects are secondary classified, and the classification result is more matched with the cause of the equipment failure or the equipment maintenance requirement, so that this classification result is used as the quality feature data of the morphological dimension, which is conducive to improving the accuracy of the alarm.

[0073] B. Time Dimension

[0074] If the quality data includes the results of defect detection on product images using a computer vision algorithm, the frequency and / or density of defects in the first preset time window can be counted, and the obtained frequency and / or density can be used as the quality feature data in the time dimension. Optionally, frequency can be defined as the number of times a defect occurs per unit time, and density can be defined as the number of consecutive times a defect occurs. Of course, other definitions are not excluded. Note that if multiple defects appear on a product, the number of times the defect occurs can be counted as 1, indicating that the product has defects, or it can be counted as the actual number of defects that occur, or there can be other statistical methods.

[0075] For example, the first preset time window is half an hour. The defect detection results of products produced within half an hour are counted. If 5 products have defects, the defect frequency is 10 pieces / hour. The corresponding alarm rules are as follows: if the defect frequency is greater than 8 pieces / hour, the industrial equipment has a fault b1.

[0076] For another example, the first preset time window is a batch of products. The defect detection results of a batch of products are counted. If a maximum of 20 continuously produced products have defects, the defect density is 20 (or the maximum value of the defect density is 20). The corresponding alarm rules are, for example: if the defect density is greater than 15, then there is a fault b2 in the industrial equipment.

[0077] According to requirements, the first preset time window can also be multiple values ​​instead of just one value. For example, the first preset time window can be 5 minutes and half an hour, indicating that defect statistics must be performed every 5 minutes and every 30 minutes, and the statistical results for 30 minutes can be obtained based on the statistical results for 5 minutes.

[0078] It should also be noted that the above examples do not distinguish between defect categories. During implementation, the frequency and / or density can be counted according to the defect category, and it is not necessary to mix all categories of defects together for statistics.

[0079] Product defects caused by equipment failure are likely not accidental, but densely distributed in the time dimension. The two indicators of frequency and density are both a reflection of the density of defects in the time dimension. Therefore, in the above implementation method, the frequency and / or density of product defects are counted, and such statistical results are used as quality feature data in the time dimension, which is conducive to improving the accuracy of alarms.

[0080] If the quality data includes measured numerical data that characterizes the performance or specifications of the product, the statistical indicators of the numerical data in the second preset time window can be counted, and the obtained statistical indicators are used as the quality characteristic data of the time dimension. Among them, the statistical indicators can depend on the business needs of the user, for example, they can be some indicators that characterize the stability of product quality, such as one or more indicators such as mean, variance, maximum value, minimum value, quantile, skewness, and kurtosis. The values ​​of the second preset time window and the first preset time window can be the same or different.

[0081] For example, the second preset time window is one hour, and the resistance values ​​of products produced within one hour are counted, and the average value is 5.2Ω, which can be used as the quality characteristic data of the time dimension. The corresponding alarm rule is, for example: if the resistance value of the product deviates from the design value of 5Ω by more than 2%, then the industrial equipment has a fault b3.

[0082] Single numerical data is accidental and may not effectively reflect equipment failures or maintenance requirements. The statistical results of numerical data within a time window are more valuable for reference in this regard. Therefore, in the above implementation method, the statistical indicators of numerical data are counted in the time dimension, and such statistical indicators are used as quality feature data in the time dimension, which is conducive to improving the accuracy of alarms.

[0083] C. Spatial Dimension

[0084] If the quality data includes the detection results of defect detection on product images using computer vision algorithms, the correlation between the defect and the equipment component can be calculated based on the defect location in the detection results and the location of the equipment component in the industrial equipment, and the obtained correlation can be used as quality feature data in the spatial dimension.

[0085] The position of the defect in the detection result may refer to the position of the detection frame containing the defect in the product image. After the industrial equipment is installed, the position of the equipment components is known, and the position can also be mapped to the image plane, so that the position of the defect and the position of the equipment components are aligned in a coordinate system, which is convenient for subsequent calculations. Of course, position alignment can also be achieved in other ways, such as aligning the two in the world coordinate system.

[0086] There are many ways to calculate the correlation between defects and equipment parts. For example, the area overlap between the defect and the equipment part can be calculated based on the defect location and the equipment part location, and the obtained overlap can be used as the correlation between the two. The area of ​​the defect can refer to the area of ​​the detection frame in which it is located. Of course, if the shape of the defect in the detection frame is further accurately identified, the defect area can be calculated more accurately; the area of ​​the equipment part can refer to the area of ​​its projection shape on the image plane, or the area of ​​the circumscribed rectangle of the projection, etc. There are also different ways to define the area overlap between the two. For example, the intersection over union (IoU) of the two areas can be used as the area overlap. The location of the defect and the equipment part is required when calculating the IoU.

[0087] For example, the area overlaps between the defect and equipment components c1, c2, and c3 are 0.05, 0.1, and 0.35, respectively, that is, the area overlap with c3 is the largest. Combined with other aspects of the alarm rule matching, it is determined whether the defect is caused by a failure in c3.

[0088] For example, the area overlaps between defects and equipment components c4, c5, and c6 are 0.04, 0.22, and 0.39, respectively, and the area overlap threshold is 0.2, which means that only those with an area overlap greater than the threshold are considered likely to cause defects. Thus, only c5 and c6 may cause defects. Combined with the matching of other alarm rules, it is determined whether the defect is caused by the failure of c5 or c6.

[0089] Since the area overlap between defects and equipment components fully reflects the degree of correlation between the two in spatial position, it is reasonable to use the area overlap as the correlation between defects and equipment components. In addition, both the defect area and the equipment component area are relatively easy to obtain, which helps to reduce the difficulty of calculating the correlation.

[0090] In addition to the area overlap, the position offset between the defect position and the position of the equipment component can also be calculated based on the defect position and the position of the equipment component, and the offset can be used as the correlation between the two. Optionally, to simplify the calculation, the defect position can be represented by one coordinate, such as the center coordinate of the detection frame, and the position of the equipment component can be represented by one coordinate, such as the center coordinate of the circumscribed rectangle of the projection shape of the equipment component on the image plane. Of course, one set of coordinates can also be used to represent the defect position and another set of coordinates can be used to represent the position of the equipment component. Each set of coordinates can include one or more coordinates, and the position offset between the two is defined as the distance between the two sets of coordinates.

[0091] Alternatively, the area overlap and position offset of the defect and the equipment component may be calculated simultaneously, and these two types of information may be combined as the correlation between the defect and the equipment component.

[0092] Alternatively, other methods are used to calculate the correlation between defects and equipment components, which will not be given as examples.

[0093] Some product defects caused by equipment failure are closely related to the location of equipment components. For example, the equipment component that causes the defect (most likely the equipment component that fails) is likely to be located near the defect. Therefore, in the above implementation method, the correlation between the defect and the equipment component is calculated using the spatial position, and this correlation is used as the quality feature data of the spatial dimension, which is conducive to improving the accuracy of the alarm.

[0094] Alternatively, the spatial dimension quality feature data may not be extracted based on the position of the equipment component. For example, the specific position of the defect in the product may be determined based on the defect position in the detection result and the product image, and the obtained specific position may be used as the spatial dimension quality feature data.

[0095] The defect position in the detection result may refer to the position of the detection frame containing the defect in the product image, which may be a set of coordinates in the image coordinate system. The specific position of the defect in the product is a specific part of the product, such as the main grid line position of the battery cell, the lower frame position, the upper left corner position, etc. The position of the product in the product image can be detected, and the product pattern is also known, so it is possible to determine which part of the product a certain position in the product image corresponds to.

[0096] For example, the defect position is (100, 200) (center coordinates of the detection frame), which corresponds to the lower border of the battery cell, and the equipment component for processing the lower border of the battery cell is known, so it can be determined that the equipment component has a fault.

[0097] Different positions of the product may be processed by different equipment components. Therefore, defects located at different positions of the product in the product image may correspond to failures of different equipment components. Therefore, in the above implementation method, the defect position in the detection result is further located to the specific position on the product in combination with the product image, and this specific position is used as the quality feature data of the spatial dimension, which is conducive to improving the accuracy of the alarm.

[0098] In addition to defect detection results and numerical data characterizing the performance or specifications of the product, quality data also includes other data. You can also refer to the above description and extract features from at least one dimension of form, time and space to obtain corresponding quality feature data.

[0099] Based on the above embodiment, optionally, the first original data may also include process data of the industrial equipment and / or operation status data of the industrial equipment.

[0100] Among them, process data refers to the process parameters used by the equipment to process the product, such as heating temperature, heating time, and the gas pressure value used. These data can be collected during the production process of the product, for example, through sensors built into the equipment or external sensors. Process data can be time series data, that is, the collected process data can be timestamped.

[0101] Operation status data refers to parameters that characterize the operation status of the device itself, such as the device's own temperature, motor vibration data, motor torque data, noise data emitted by the device itself, etc. These data can be collected during the product production process, for example, through sensors built into the device or external sensors. Operation status data can be time series data or non-time series data.

[0102] The sensors mentioned above are selected according to the data collection requirements, such as temperature sensors, pressure sensors, vibration sensors, sound sensors, etc. For the sensors that come with industrial equipment, the industrial equipment can also transmit the real-time data collected by these sensors to the Programmable Logic Controller (PLC) through the input port of the PLC, and the server (execution alarm method) then connects to the PLC to obtain these data. Of course, the solution of industrial equipment directly connecting to the server is not ruled out. For external sensors, they can also be directly connected to the server without being transferred through the PLC. Of course, the solution of transferring through the PLC is not ruled out.

[0103] If the first original data also includes process data of industrial equipment and / or operating status data of industrial equipment, the process data and / or operating status data may also be processed in step S120. The processing method may include preprocessing, feature extraction of different dimensions, etc. For details, please refer to the processing method of quality data, which will not be repeated here.

[0104] In the above implementation, the first original data may further include data related to the device, which enriches the data source of the alarm analysis and is therefore helpful to improve the accuracy or comprehensiveness of the alarm.

[0105] Based on the above embodiment, optionally, the first original data may also include incoming material data during the production of industrial equipment.

[0106] Incoming material data refers to the upstream incoming materials and auxiliary material data used in the production of industrial equipment, which may include one or more data such as the name, manufacturer, batch, specification, and quality of the incoming materials and auxiliary materials.

[0107] For example, more defects were detected in the products produced by device M1, while fewer defects were detected in the products produced by device M2 of the same model. The incoming material data of devices M1 and M2 can be obtained, and the alarm rule can be: if the incoming materials of devices M1 and M2 are the same, then device M2 has a fault or needs maintenance (of course, it can also be judged in combination with other information).

[0108] For another example, equipment M1 produces 10 batches of products, some of which use raw material Q1, and the remaining batches use raw material Q2. The alarm rule can be: if the frequency of defects in the batches using raw material Q1 is significantly higher (for example, 3 times or more) than the frequency of defects in the batches using raw material Q2, it means that there is a problem with raw material Q1 and it needs to be replaced (of course, it can also be judged in combination with other information). Strictly speaking, the prompt to replace raw materials does not belong to equipment fault information or equipment maintenance information, but because raw materials are used by equipment to produce products and are closely related to equipment and products, it is also acceptable to classify them as fault information or maintenance information for unified analysis and processing.

[0109] For another example, auxiliary material data may include not only auxiliary materials but also data of some equipment components (such as some components that need to be replaced regularly), and the failure or maintenance of these equipment components may be involved in the alarm rules.

[0110] For another example, the incoming material data may include the battery cell pattern. As described above, the battery cell pattern can be used to determine the specific location of the defect in the product, and then extract the quality feature data in the spatial dimension.

[0111] According to the above example, the incoming material data may be used directly to match the alarm rules, or may be used to form alarm rules, extract features of quality data, and other purposes. Since the incoming material data directly reflects the characteristics of the product itself, it is helpful to improve the accuracy or comprehensiveness of the alarm.

[0112] In addition to product quality data, industrial equipment process data, industrial equipment operating status data, and industrial equipment incoming material data, some other data can also be collected for alarm analysis, although these data may not be directly used to match alarm rules. Here are some such data:

[0113] (1) Historical alarm events of industrial equipment and their handling methods

[0114] For example, it may include the time when the alarm event occurred, the data that caused the alarm, and the specific content of the alarm information (including solutions). These data can be used to construct alarm rules.

[0115] (2) Process formula data

[0116] Process formula data includes: the raw materials, auxiliary materials and their usage (formula data) required to produce a certain product during the product manufacturing process, as well as specific process operation steps and parameter settings (process data). Process formula data is used to ensure the stability of the production process and the quality of the final product, covering all key links from the use of raw materials to the production of finished products. For example, for different formulas, the quality data of the product will be different, so the parameters in the alarm rules need to be adjusted accordingly.

[0117] (3) Location data of equipment components

[0118] It can be used to calculate the quality characteristic data of the spatial dimension, as described above.

[0119] Note that the process formula data may contain product raw material data, but the information is relatively rough, mainly including the name and amount of the raw materials, while the incoming material data may include more detailed raw material information. In addition, the formula data may contain the process parameters of the equipment, such as heating temperature, but the process parameters in the formula data are set values, not real-time collected values, and therefore are different from the process data of the equipment.

[0120] Below, based on the above embodiments, some examples are given to illustrate the application of some alarm rules:

[0121] Rule 1: For photovoltaic modules processed by the stack welding machine, the abnormal shape of the welding rod is detected through visual defects (visual quality data). On this basis, it is analyzed that the abnormal welding rod appears at the head and tail of the photovoltaic module (quality feature data in the spatial dimension). Further, according to the alarm rule, this defect is associated with the problem of string collision in the typesetting machine. Among them, the typesetting machine is located before the stack welding machine on the production line of photovoltaic modules, and is mainly used to typeset the battery strings into photovoltaic modules.

[0122] Rule 2: For photovoltaic modules processed by the stitch welding machine, the computer vision algorithm detects the presence of secondary tin solder joints on the cells (visual quality data). Based on this, it is analyzed that the location of the solder joint is at the welding head (quality feature data in the spatial dimension). Further, according to the alarm rules, this defect is associated with problems in the welding time, current or cooling time of the stitch welding machine.

[0123] Rule 3: For photovoltaic modules processed by the stack welding machine, the computer vision algorithm detects that there is poor spacing between the solar cell modules (visual quality data), and further associates this defect with deviation correction abnormality according to the alarm rule. Deflection correction abnormality refers to the abnormal layout of photovoltaic modules by the stack welding machine. The possible cause of the fault is that there is a problem with the correction force of the front glass of the stack welding machine.

[0124] Rule 4: It is detected that the device torque (numerical quality data) of the slicer exceeds a certain threshold. This situation is further associated with the presence of foreign matter in the broken wire net (an auxiliary material of the slicer) based on the alarm rule, and cleaning and maintenance are required.

[0125] It should be noted that some faults or maintenance requirements are more complex, and the corresponding alarm rules may involve multiple first-order data to be analyzed at the same time. For example, based on the quality characteristic data of the spatial dimension, it is determined that a certain defect appears in the center of the product, and then combined with the quality characteristic data of the time dimension to determine that the defect appears continuously on multiple products, it is finally determined that a certain fault has occurred in the industrial equipment.

[0126] Based on the above embodiments, optionally, Figure 1 The alarm methods also include:

[0127] Step S140: receiving first user feedback for the first alarm information, and iteratively optimizing the alarm rules in the rule base according to the first user feedback.

[0128] The first user feedback includes feedback on the authenticity of the first alarm information, and may include, but is not limited to, at least one of the following feedback:

[0129] (1) Whether the fault in the first alarm information actually exists;

[0130] (2) Whether the maintenance requirement in the first alarm information actually exists;

[0131] (3) Whether the fault cause in the first fault alarm information is correct. If it is not correct, the correct fault cause can be fed back;

[0132] (4) Whether the maintenance suggestion in the first fault alarm information is correct. If not, the correct maintenance suggestion can be fed back.

[0133] Users can fill in feedback information through a system interface (such as a web page or client interface) and submit it. Of course, other methods can also be used, such as users directly speaking or inputting feedback information in natural language, and the system analyzing the content of the user's feedback through a large language model, etc.

[0134] The alarm rules in the rule base include a series of parameters (for example, thresholds), which can be called rule parameters. Iterative optimization is to continuously optimize the values ​​of these rule parameters through some algorithms, so that the accuracy of the alarm rules is getting higher and higher. For example, the iterative optimization of rule parameters can be performed through the Bayesian optimization algorithm, which specifically includes the following steps:

[0135] Step 1: Initialize the value combination of rule parameters

[0136] For example, the rule parameters can be initialized to some empirical values.

[0137] Step 2: Evaluate the value combination of rule parameters

[0138] According to the first original data obtained, an alarm is issued for the first original data based on the value combination of the rule parameters, a first alarm message is output, and first user feedback on the first alarm message is collected, and the alarm effect of the value combination of the rule parameters is evaluated using the feedback.

[0139] Step 3: Build a proxy model and predict the alarm accuracy

[0140] According to the evaluation result of step 2, Gaussian Process Regression (GPR) in the Bayesian optimization algorithm is used to model the relationship between the rule parameters and the alarm accuracy to obtain the proxy model. Optionally, if the evaluation result of step 2 is good, and a proxy model has been built when step 3 is executed, the proxy model may not be rebuilt, but only some adjustments may be made to the parameters of the proxy model; if the evaluation result of step 2 is poor, even if a proxy model has been built when step 3 is executed, the proxy model may be rebuilt.

[0141] The proxy model is used to predict the alarm accuracy corresponding to the value combination of rule parameters.

[0142] Step 4: Determine the new value combination of the rule parameters

[0143] The Bayesian optimization algorithm determines the next value combination to be tried through the acquisition function. The acquisition function can find the current optimal value combination based on the alarm accuracy and uncertainty predicted in step 3.

[0144] Jump to step 2 to continue execution. Note that the new value combination obtained in step 4 before the jump is the value combination in step 2 after the jump.

[0145] The above iterative optimization steps may be performed for each rule separately, or may be performed for several rules together.

[0146] Iterative optimization can be performed continuously as the product is produced, or optimization can be stopped after the rule parameters are optimized to a certain extent (for example, the evaluation result in step 2 is better than a certain standard), and so on.

[0147] In addition, some rule parameters may be given by independent model predictions. During iterative optimization, the parameters of these models can also be optimized so that the optimized models can give more accurate rule parameters.

[0148] In the above implementation, the alarm rules in the rule base can be iteratively optimized based on user feedback, which is conducive to integrating more manual experience into the rules and improving the accuracy of alarms.

[0149] Figure 2 A flowchart of another alarm method for industrial equipment provided in an embodiment of the present application. The method can be, but is not limited to, Figure 4 The electronic device shown is executed, and the structure of the electronic device can be referred to later. Figure 4 's explanation. Figure 1 The method (including its various possible implementation methods) is mainly aimed at warnings after the product is produced, because the method needs to use the quality data of the product. Figure 2 The method (including its various possible implementation methods) is mainly aimed at alarms during the product production process. The method does not need to use product quality data because the product may still be in production at this time. These two methods can be implemented one by one or at the same time. Figure 2 , the alarm methods for industrial equipment include:

[0150] Step S210: Acquire second original data during the operation of the industrial equipment.

[0151] The second original data includes the process data of the industrial equipment and / or the operating status data of the industrial equipment. As can be seen from the foregoing, the process data of the industrial equipment and / or the operating status data of the industrial equipment can also be used as the first original data, and the method of obtaining the data has been explained, which will not be repeated here. If step S210 is executed first, and then step S110 is executed, and the first original data includes the process data of the industrial equipment and / or the operating status data of the industrial equipment, then the data obtained in step S210 can be used directly without repeated acquisition.

[0152] It should be noted that since the second raw data is obtained during the operation of industrial equipment, Figure 2 The method can be executed in real time during the operation of the industrial equipment. Of course, the method can also be implemented in non-real time, such as first saving the collected second original data, and then performing alarm analysis based on the second original data at a suitable time in the future.

[0153] Step S220: obtaining second data to be analyzed according to the second original data.

[0154] Step S220 may refer to step S120 and will not be described again.

[0155] Step S230: using the first anomaly detection model to perform anomaly detection on the second data to be analyzed to obtain anomaly detection results.

[0156] The input of the first anomaly detection model may include the second data to be analyzed, and the output may include an anomaly detection result. The detection result may be whether an anomaly exists in the second data to be analyzed (which may also be regarded as a binary classification of the anomaly), or a more detailed anomaly classification may be given.

[0157] The first anomaly detection model can be a neural network model, a decision tree model, a support vector machine model, etc., without limitation. The first anomaly detection model can be a single model or include multiple sub-models. If the first anomaly detection model includes multiple sub-models, these sub-models can perform anomaly detection on the second data to be analyzed respectively to obtain corresponding anomaly detection sub-results, and then use a certain integration method to obtain the final anomaly detection result based on these anomaly detection sub-results. Obtaining the final anomaly detection result by integrating the anomaly detection sub-results of each sub-model is conducive to learning from each other's strengths and weaknesses, thereby improving the robustness of the detection result.

[0158] The following examples illustrate possible integration methods:

[0159] (1) Hard voting

[0160] Count the anomaly categories in all anomaly detection sub-results, and take the anomaly category with the most votes as the final anomaly detection result. For example, if the first anomaly detection model includes three sub-models, namely the isolation forest model, the single-class support vector machine model, and the autoencoder model, these three sub-models can be used to perform anomaly detection on the second data to be analyzed, and the detection results are as follows:

[0161] Isolation forest model: The probability of anomaly in class X is 0.6, the probability of anomaly in class Y is 0.2, and the probability of anomaly in class Z is 0.2; the converted vote is 1 vote for anomaly in class X.

[0162] Single-class support vector machine model: the probability of anomaly of class X is 0.5, the probability of anomaly of class Y is 0.4, and the probability of anomaly of class Z is 0.1; the converted vote is 1 vote for anomaly of class X.

[0163] Autoencoder model: The probability of anomaly of class X is 0.45, the probability of anomaly of class Y is 0.5, and the probability of anomaly of class Z is 0.05; the converted vote is 1 vote of anomaly of class Y.

[0164] According to the above three voting results, the X-type anomaly has 2 votes, the Y-type anomaly has 1 vote, and the Z-type anomaly has 0 votes. Since the X-type anomaly has the most votes, the final anomaly detection result is the X-type anomaly.

[0165] (2) Soft voting

[0166] Perform weighted calculation on multiple anomaly detection sub-results and determine the final anomaly detection result based on the weighted calculation result. Using the example in (1):

[0167] Isolation forest model: X-class anomaly probability 0.6, Y-class anomaly probability 0.2, Z-class anomaly probability 0.2; weight is 0.3.

[0168] Single-class support vector machine model: the anomaly probability of class X is 0.5, the anomaly probability of class Y is 0.4, the anomaly probability of class Z is 0.1; the weight is 0.3.

[0169] Autoencoder model: X-class anomaly probability 0.45, Y-class anomaly probability 0.5, Z-class anomaly probability 0.05; weight is 0.4.

[0170] The probability of each type of anomaly is weighted:

[0171] Probability of abnormality of type X = 0.6 × 0.3 + 0.5 × 0.3 + 0.45 × 0.4 = 0.51;

[0172] Probability of abnormality of type Y = 0.2 × 0.3 + 0.4 × 0.3 + 0.5 × 0.4 = 0.38;

[0173] Probability of class Z anomaly = 0.2 × 0.3 + 0.1 × 0.3 + 0.05 × 0.4 = 0.11;

[0174] Since the probability of class X anomalies is the highest, the final anomaly detection result is class X anomalies.

[0175] There are different ways to determine the weights used in the above weighted calculations. For example, they can be determined directly based on experience, or they can be determined based on the performance evaluation results of each sub-model. For example, the weights can be positively correlated with the performance of the sub-model. The performance here can be defined as the accuracy of the sub-model in detecting anomalies. The performance evaluation results of the sub-model can be obtained by testing the sub-model using some test data, or they can be determined directly based on the structural characteristics of the sub-model (for example, for similar models, it can be simply considered that the model performance is positively correlated with the number of model layers), or they can be determined by combining the above two evaluation methods, and so on.

[0176] The weights of the anomaly detection sub-results for weighted calculation are determined based on the performance evaluation results of the sub-models, which fully considers the impact of model performance differences on the accuracy of anomaly detection, thereby facilitating more accurate anomaly detection results and thus improving the accuracy of alarms.

[0177] A corresponding first anomaly detection model may be set for each type of second data to be analyzed, or multiple types of second data to be analyzed may share the same first anomaly detection model.

[0178] The first anomaly detection model is a model obtained by unsupervised training. Unsupervised training is training using unlabeled training samples. The labels here can be understood as the actual anomaly detection results corresponding to the training samples, such as the actual anomaly categories.

[0179] Step S240: If the abnormality detection result indicates that the second data to be analyzed is abnormal, a second alarm message is output based on the abnormality detection result.

[0180] The second warning information may include at least one of the following:

[0181] (1) Data anomaly information: for example, whether there are any anomalies in the second data to be analyzed, and the specific anomaly category.

[0182] (2) Fault information of industrial equipment associated with data anomalies: The fault information has been explained in the previous article and will not be repeated here. In order to output the information in item (2), it is necessary to first establish an association between the data anomaly and the equipment fault information, or to form a corresponding mapping rule. For example, temperature anomaly may be due to a bearing failure in the equipment, motor amplitude anomaly may be due to the looseness of some mechanical parts, and so on. The above rules can be fixed or optimized over time (including addition, modification, and deletion). For example, the anomaly of a certain data was not associated with the equipment failure at the beginning. After the anomaly of the data occurred repeatedly, it was manually analyzed and the anomaly of the data was associated with the equipment failure.

[0183] (3) Maintenance information of industrial equipment associated with data anomalies: The maintenance information has been explained in the previous article and will not be repeated here. In order to output the information in item (3), it is necessary to first establish an association between the data anomaly and the equipment maintenance information, or to form a corresponding mapping rule. For example, temperature anomalies require maintenance of the equipment's bearings. The above rules can be fixed or optimized over time (including additions, modifications, and deletions). For example, when explaining the information in item (2), it was mentioned that temperature anomalies are considered bearing failures. Based on this rule, further rules for bearing maintenance can be added to try to provide early warnings and reduce the situation where alarms are only issued when the bearing is actually faulty. However, there may be certain differences in the judgment of temperature anomalies for maintenance and temperature anomalies for faults. For example, for the latter, the temperature may be higher before it is judged as an abnormality.

[0184] There are many ways to output the second warning information, which can refer to the output method of the first warning information, and will not be repeated. If the abnormality detection result shows that the second data to be analyzed does not have an abnormality, no processing is required.

[0185] Brief summary Figure 2 The method in supports anomaly detection based on the acquired second raw data during the operation of industrial equipment, so as to timely discover anomalies in the data and issue an alarm. Since data anomalies may be caused by equipment failure or may be a sign that the equipment needs maintenance, this implementation method is conducive to early detection of equipment failure or maintenance of the equipment (rather than waiting until the product is produced). In addition, anomaly detection uses the first anomaly detection model obtained by unsupervised training, that is, there is no need to collect labeled training samples in advance, which is conducive to performing this anomaly detection in the early stage of production line operation and in the absence of data accumulation, and can also detect equipment failure or maintain the equipment early.

[0186] Continue to refer to Figure 2 , optional, Figure 2 The alarm method may also include:

[0187] Step S250: receiving second user feedback for the second alarm information, and constructing a labeled training sample according to the second user feedback and the second data to be analyzed.

[0188] The second user feedback includes feedback on the authenticity of the second alarm information, and may include, but is not limited to, at least one of the following feedback:

[0189] (1) Whether the data anomaly in the second alarm information actually exists;

[0190] (2) Whether the data anomaly category in the second alarm information is correct. If it is not correct, the correct anomaly category can be fed back;

[0191] (3) Whether the fault in the second alarm information actually exists;

[0192] (4) Whether the maintenance requirement in the second alarm information actually exists;

[0193] (5) Whether the fault cause in the second fault alarm information is correct. If it is not correct, the correct fault cause can be fed back;

[0194] (6) Check whether the maintenance suggestion in the second fault alarm information is correct. If not, the correct maintenance suggestion can be fed back.

[0195] The method for obtaining the second user feedback may refer to the method for obtaining the first user feedback, and will not be elaborated again.

[0196] The label of the second data to be analyzed, such as a real abnormal category label, can be determined according to the second user feedback, and the second data to be analyzed with the label can be used as a training sample in step S260.

[0197] Step S260: When the number of collected training samples meets the training requirements, supervised training is performed using the training samples to obtain a second anomaly detection model.

[0198] The training requirements are set by the user. For example, if the collected training samples exceed the preset number, model training will be started, or if the time for collecting training samples exceeds the preset duration, model training will be started, or if no new labeled training samples are collected within a preset period of time, model training will be started based on the already collected training samples. Of course, model training can also be performed once as long as labeled training samples are collected, and so on.

[0199] The second anomaly detection model is similar to the first anomaly detection model in function. For example, it can be a single model or include multiple sub-models. A corresponding second anomaly detection model can be set for each type of third data to be analyzed, or multiple types of third data to be analyzed can share the same second anomaly detection model. If there are multiple second anomaly detection models, and different second anomaly detection models target different types of third data to be analyzed, then these second anomaly detection models may not necessarily perform supervised training synchronously, because the collection progress of training samples corresponding to different types of third data to be analyzed may not be the same, and the requirements of each second anomaly detection model for the number of training samples may not be the same.

[0200] In addition, the second anomaly detection model may be the same as the first anomaly detection model in terms of model selection (for example, both are decision tree models), or it may be different from the first anomaly detection model. For example, the second anomaly detection model includes two sub-models, namely a random forest model and a long short-term memory network, which are different from the three sub-models that may be included in the first anomaly detection model mentioned above.

[0201] Optionally, after the second anomaly detection model is trained and put into use, labeled training samples may continue to be collected, and these training samples may be used to continue iteratively optimize the second anomaly detection model.

[0202] Step S270: Acquire the third original data.

[0203] Step S280: Obtain third data to be analyzed according to the third original data.

[0204] Steps S270 and S280 are basically the same as steps S210 and S220, and will not be described in detail. The third original data is actually the second original data, but it is the second original data used by the second anomaly detection model after the second anomaly detection model is trained. In order to distinguish it from the second original data before the second anomaly detection model is put into use, it is renamed.

[0205] Steps S290 and S292 are two parallel solutions, and you can choose to implement either one. Among them, step S290 directly replaces the first anomaly detection model with the second anomaly detection model, and the first anomaly detection model is no longer used for anomaly detection in the future. Step S292 continues to retain the first anomaly detection model, and uses the first anomaly detection model and the second anomaly detection model for anomaly detection at the same time.

[0206] Step S290: using the second anomaly detection model, perform anomaly detection on the third data to be analyzed to obtain anomaly detection results.

[0207] Step S290 is similar to step S230 and will not be described again.

[0208] Step S292: Combine the second anomaly detection model and the first anomaly detection model to perform anomaly detection on the third data to be analyzed to obtain anomaly detection results.

[0209] There are many ways to combine the first anomaly detection model and the second anomaly detection model for detection. For example, the detection results of the two can be integrated, or the first anomaly detection model or the second anomaly detection model can be used to perform anomaly detection under different conditions, or the first anomaly detection model and the second anomaly detection model can be used in chronological order to perform anomaly detection, etc. More specific examples will be given later.

[0210] In steps S250-S292, a sufficient number of labeled training samples can be constructed based on user feedback, so as to perform supervised training to obtain a second anomaly detection model. Since the second anomaly detection model absorbs more manual experience, it is likely to be better than the first anomaly detection model obtained based on unsupervised training in terms of performance. Therefore, when performing anomaly detection on the subsequent third original data, the first anomaly detection model can be replaced by the second anomaly detection model to improve the accuracy of anomaly detection. Of course, the first anomaly detection model can also be retained, and the two anomaly detection models can be combined for anomaly detection to obtain more reliable anomaly detection results. For example, for some accidental anomalies, the second anomaly detection model that performs supervised training may perform poorly because they are not included in the labeled training samples. Instead, the first anomaly detection model that does not rely on labels performs better.

[0211] Based on the above embodiments, the following further introduces possible combinations of the two anomaly detection models in step S292:

[0212] Combination method 1

[0213] First, the second anomaly detection model is used to perform anomaly detection on the third data to be analyzed to obtain a preliminary anomaly detection result. If the preliminary anomaly detection result shows that the third data to be analyzed does not have an anomaly, the first anomaly detection model is further used to perform anomaly detection on the third data to be analyzed to obtain a final anomaly detection result; if the preliminary anomaly detection result shows that the third data to be analyzed has an anomaly, the preliminary anomaly detection result is directly determined as the final anomaly detection result.

[0214] Combined with method 1, the second anomaly detection model is used preferentially for anomaly detection. Since the second anomaly detection model is obtained through supervised training, it has a high detection accuracy in most cases. If the second anomaly detection model fails to detect an anomaly, it does not necessarily mean that there is no anomaly. It may just be that the training samples of the second anomaly detection model do not contain anomalies with such labels, resulting in failure to detect them. At this time, the first anomaly detection model obtained through unsupervised training is used for detection. Since the first anomaly detection model has a low reliance on labels during training, it is possible to discover missed anomalies, thereby improving the accuracy of anomaly detection.

[0215] Combination method 2

[0216] First, the first anomaly detection model and the second anomaly detection model are used to perform anomaly detection on the third data to be analyzed. The detection process can be performed serially or in parallel. The detection result obtained by the former is the first anomaly detection result, and the detection result obtained by the latter is the second anomaly detection result. Then, the final anomaly detection result is obtained based on the first anomaly detection result and the second anomaly detection result. For example, the soft voting method introduced above can be used to perform weighted calculation on the first anomaly detection result and the second anomaly detection result to obtain the final anomaly detection result.

[0217] Combination method 2 obtains the final anomaly detection result by integrating the detection results of the first anomaly detection model and the second anomaly detection model, which is equivalent to combining the advantages of supervised learning and unsupervised learning, thereby helping to improve the accuracy of anomaly detection.

[0218] Furthermore, if soft voting is used to implement combination method 2, the weights used in the weighted calculation are dynamically adjusted based on at least one of the following factors:

[0219] (1) Data quality of the third raw data

[0220] For example, when the data quality of the third original data is poor, the weight of the first anomaly detection model can be increased, while the weight of the second anomaly detection model can be reduced (the sum of the weights of the two models can be a fixed value, such as 1); when the data quality of the third original data is good, the weight of the first anomaly detection model can be reduced, while the weight of the second anomaly detection model can be increased. The principle is that when the data quality is poor, the characteristics of the abnormal data are relatively insignificant, so that more reliance can be placed on the detection results of the unsupervised model; when the data quality is good, the characteristics of the abnormal data are relatively significant, so that more reliance can be placed on the detection results of the supervised model.

[0221] As for the data quality of the third original data, it can be judged by using statistical methods, artificial intelligence methods or some rules. For example, if there are many missing values ​​in the third original data, it is judged that the data quality is poor.

[0222] (2) Data type of the third original data

[0223] For example, for temperature data, the first anomaly detection model performs better in detecting anomalies, so its weight can be increased, while the weight of the second anomaly detection model can be reduced; for pressure data, the first anomaly detection model performs poorly in detecting anomalies, so its weight can be reduced, while the weight of the second anomaly detection model can be increased.

[0224] (3) Performance evaluation results of the first anomaly detection model and the second anomaly detection model

[0225] For example, when the second anomaly detection model is first put into use, the first anomaly detection model has better performance and a larger weight, while the second anomaly detection model has a smaller weight. As the second anomaly detection model is continuously optimized, its performance gets better and better, and the corresponding weight also increases.

[0226] As for how to evaluate model performance, we have already introduced it in the previous article when introducing the sub-model of the first anomaly detection model, so we will not repeat it here.

[0227] (4) Time lapse

[0228] For example, when the second anomaly detection model is first put into use, the first anomaly detection model has better performance and a larger weight, while the weight of the second anomaly detection model is smaller. As time goes by, the weight of the second anomaly detection model gradually increases, and the weight of the first anomaly detection model gradually decreases.

[0229] The principle of item (4) is similar to that of item (3), except that the passage of time and the improvement of model performance are not necessarily completely synchronized. When it is difficult to determine the model performance, it can be assumed that the performance of the second anomaly detection model will continue to improve over time (as long as labeled training samples are continuously collected).

[0230] (5) Working scenarios of industrial equipment

[0231] The working scenario here can represent the busyness of the equipment or the external environment, etc. For example, when the industrial equipment is in a high-load scenario, the vibration pattern of its motor is different from usual, resulting in a decrease in the performance of the second abnormality detection model, so its weight can be reduced at this time, while the weight of the first abnormality detection model can be increased; when the industrial equipment is in a medium-to-low load scenario, its motor vibration pattern is the same as usual, and the performance of the second abnormality detection model is better, so its weight can be increased at this time, while the weight of the first abnormality detection model can be reduced.

[0232] (6) Maintenance status of industrial equipment

[0233] For example, there is a big difference in the status of industrial equipment before and after maintenance. This difference will be reflected in the data related to the equipment, which will affect the results of anomaly detection. Therefore, after each maintenance, the weights of the first anomaly detection model and the second anomaly detection model can be adjusted.

[0234] The above six weight determination factors can also be combined with each other to form a more refined weight determination rule when there is no contradiction, and no more examples are given. Based on these factors, the weights of the weighted calculation of the first anomaly detection result and the second anomaly detection result can be dynamically adjusted, which is conducive to flexible setting of weights in different scenarios, giving full play to the advantages of supervised learning and unsupervised learning in different scenarios, so as to obtain more accurate anomaly detection results.

[0235] In addition, other methods of combining the first anomaly detection model and the second anomaly detection model for anomaly detection are not completely excluded. For example, the first anomaly detection model is first used to detect the third data to be analyzed, and then whether to use the second anomaly detection model for detection is decided based on the detection results, and so on.

[0236] Combination Figure 1 and Figure 2 The method in (including its various possible implementation methods), the alarm method for industrial equipment provided in the embodiment of the present application can provide fault and / or maintenance alarms for industrial equipment in various links such as before, during and after product production, thereby helping to comprehensively improve the operational stability of the industrial assembly line and thereby improve product quality.

[0237] Figure 3 This is a module diagram of an alarm device 300 for industrial equipment provided in an embodiment of the present application. Figure 3 , an alarm device 300 for industrial equipment (hereinafter sometimes referred to as the alarm device 300) includes:

[0238] The data acquisition module 310 is used to acquire first original data, where the first original data includes quality data of products produced by the industrial equipment;

[0239] A data processing module 320 is used to obtain first data to be analyzed according to the first original data; wherein the first data to be analyzed includes quality feature data, and the quality feature data is obtained by extracting features from the quality data;

[0240] The alarm module 330 is used to match the first data to be analyzed with the alarm rules in the rule base, and if the match is successful, output the first alarm information based on the matched alarm rules; wherein the first alarm information includes the fault information and / or maintenance information of the industrial equipment.

[0241] The alarm device 300 for industrial equipment provided in the embodiment of the present application can be used to implement the alarm method for industrial equipment provided in the embodiment of the present application. Its implementation principle and the technical effects produced have been introduced in the aforementioned method embodiment. Therefore, for the sake of brief description, for matters not mentioned in the device embodiment, please refer to the corresponding content in the method embodiment.

[0242] Figure 4 This is a structural diagram of an electronic device provided in an embodiment of the present application. Figure 4 The electronic device 400 includes: a processor 410, a memory 420 and a communication module 430. These components are interconnected and communicate with each other through a communication bus 440 and / or other forms of connection mechanisms (not shown).

[0243] The processor 410 includes one or more (only one is shown in the figure), which can be an integrated circuit chip with signal processing capability. The above-mentioned processor 410 can be a general-purpose processor, including a central processing unit (CPU), a micro control unit (MCU), a network processor (NP), or other conventional processors; it can also be a dedicated processor, including a graphics processor (GPU), a neural network processor (NPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, and discrete hardware components. In addition, when there are multiple processors 410, some of them can be general-purpose processors, and the other part can be dedicated processors.

[0244] The memory 420 includes one or more (only one is shown in the figure), which may be, but not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The processor 410 and other possible components may access the memory 420 and read and / or write data therein.

[0245] In particular, one or more computer program instructions may be stored in the memory 420 , and the processor 410 may read and execute these computer program instructions to implement the alarm method for industrial equipment provided in the embodiment of the present application.

[0246] The communication module 430 includes one or more (only one is shown in the figure), which is used for communication between the electronic device 400 and other devices, such as receiving original data from other devices. The communication module 430 can be a wired communication module or a wireless communication module.

[0247] Understandably, Figure 4 The structure shown is for illustration only. The electronic device 400 may also include Figure 4 For example, if the electronic device 400 does not need to communicate with other devices, the electronic device 400 may not include a communication module.

[0248] Figure 4 The components shown in the figure can be implemented by hardware, software or a combination thereof. The electronic device 400 may be a physical device, such as a server, a PC, a laptop, a tablet computer, a mobile phone, etc., or a virtual device, such as a virtual machine, a virtualized container, etc. Moreover, the electronic device 400 is not limited to a single device, but may also be a combination of multiple devices or a cluster consisting of a large number of devices.

[0249] The present application also provides a computer-readable storage medium on which computer program instructions are stored. When these computer program instructions are read and executed by a computer processor, the alarm method for industrial equipment provided by the present application is executed. For example, the computer-readable storage medium can be implemented as Figure 4The memory 420 in the electronic device 400, or a mobile hard disk, optical disk, USB flash disk, etc. independent of the electronic device 400.

[0250] The present application also provides a computer program product, which includes computer program instructions. When these computer program instructions are read and executed by a computer processor, the alarm method for industrial equipment provided by the present application is executed. For example, the computer program product can be stored in Figure 4 The computer program product is stored in the memory 420 in the electronic device 400 and read and executed by the processor 410. Of course, the computer program product can also be stored in a storage medium independent of the electronic device 400, such as a mobile hard disk, an optical disk, or a USB flash drive.

[0251] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An alarm method for industrial equipment, characterized in that: include: Acquire first original data, wherein the first original data includes quality data of a product produced by the industrial equipment; Obtaining first data to be analyzed according to the first original data; wherein the first data to be analyzed includes quality feature data, and the quality feature data is obtained by extracting features from the quality data, wherein at least the quality feature data of the time dimension obtained by extracting features from the quality data from the time dimension and the quality feature data of the space dimension obtained by extracting features from the quality data from the space dimension, and the quality feature data of the time dimension represents the features presented by the quality data over time; Matching the first data to be analyzed with the alarm rules in the rule base, and if the match is successful, outputting first alarm information based on the matched alarm rules; wherein the first alarm information includes fault information and / or maintenance information of the industrial equipment; If the quality data includes a detection result of defect detection on a product image of the product using a computer vision algorithm, feature extraction is performed on the quality data from a spatial dimension to obtain the quality feature data of the spatial dimension, including: According to the defect position in the detection result and the position of the equipment component in the industrial equipment, the area overlap and / or position offset of the defect and the equipment component are calculated, and the area overlap and / or the position offset are used as the correlation between the defect and the equipment component, and the correlation is used as the quality feature data in the spatial dimension.

2. The alarm method according to claim 1, characterized in that: The quality data include: The detection result of defect detection on the product image of the product using a computer vision algorithm; and / or, the measured numerical data representing the performance or specification of the product.

3. The alarm method according to claim 1, characterized in that: Extracting features from the quality data to obtain the quality feature data also includes: Feature extraction is performed on the quality data from the morphological dimension to obtain the quality feature data of the morphological dimension.

4. The alarm method according to claim 1, characterized in that: If the quality data includes a detection result of defect detection on a product image of the product using a computer vision algorithm, feature extraction is performed on the quality data from a time dimension to obtain the quality feature data of the time dimension, including: Counting the frequency and / or density of defects occurring within a first preset time window, and using the frequency and / or density as the quality characteristic data of the time dimension; If the quality data includes measured numerical data characterizing the performance or specification of the product, then feature extraction is performed on the quality data from a time dimension to obtain the quality feature data of the time dimension, including: Statistical indicators of the numerical data within the second preset time window are counted, and the statistical indicators are used as the quality feature data of the time dimension.

5. The alarm method according to claim 1, characterized in that: The first original data also includes at least one of the following: process data of the industrial equipment, operating status data of the industrial equipment, and incoming material data during the production of the industrial equipment.

6. The alarm method according to claim 1, characterized in that: The method further comprises: Receive first user feedback on the first alarm information, and iteratively optimize the alarm rules in the rule base according to the first user feedback; wherein the first user feedback includes feedback on the authenticity of the first alarm information.

7. The alarm method according to any one of claims 1 to 6, characterized in that: The method further comprises: During the operation of the industrial equipment, second original data is acquired, where the second original data includes process data of the industrial equipment and / or operation status data of the industrial equipment; Obtaining second data to be analyzed according to the second original data; Using a first anomaly detection model to perform anomaly detection on the second data to be analyzed to obtain an anomaly detection result; wherein the first anomaly detection model is a model obtained by unsupervised training; If the anomaly detection result indicates that the second data to be analyzed is abnormal, a second alarm information is output based on the anomaly detection result; wherein the second alarm information includes at least one of the following: data anomaly information, fault information of the industrial equipment associated with the data anomaly, and maintenance information of the industrial equipment associated with the data anomaly.

8. The alarm method according to claim 7, characterized in that: The first anomaly detection model includes a plurality of sub-models, and the using the first anomaly detection model to perform anomaly detection on the second data to be analyzed to obtain anomaly detection results includes: Using each sub-model to perform anomaly detection on the second data to be analyzed, respectively, to obtain corresponding anomaly detection sub-results; A weighted calculation is performed on multiple anomaly detection sub-results to obtain an anomaly detection result; wherein the weight used in the weighted calculation depends on the performance evaluation result of each sub-model.

9. The alarm method according to claim 7, characterized in that: The method further comprises: receiving second user feedback for the second alarm information, and constructing a labeled training sample according to the second user feedback and the second data to be analyzed; wherein the second user feedback includes feedback on the authenticity of the second alarm information; When the number of the collected training samples meets the training requirements, supervised training is performed using the training samples to obtain a second anomaly detection model; Acquire third original data, where the third original data is subsequent data of the second original data; Obtaining third data to be analyzed according to the third original data; Using the second anomaly detection model, anomaly detection is performed on the third data to be analyzed to obtain an anomaly detection result; or, combining the second anomaly detection model and the first anomaly detection model, anomaly detection is performed on the third data to be analyzed to obtain an anomaly detection result.

10. The alarm method according to claim 9, characterized in that: The combining the second anomaly detection model and the first anomaly detection model to perform anomaly detection on the third data to be analyzed to obtain an anomaly detection result includes: Using the second anomaly detection model to perform anomaly detection on the third data to be analyzed to obtain a preliminary anomaly detection result; If the preliminary anomaly detection result indicates that the third data to be analyzed does not have an anomaly, the first anomaly detection model is used to perform anomaly detection on the third data to be analyzed to obtain an anomaly detection result; otherwise, the preliminary anomaly detection result is directly determined as the anomaly detection result.

11. The alarm method according to claim 9, characterized in that: The combining the second anomaly detection model and the first anomaly detection model to perform anomaly detection on the third data to be analyzed to obtain an anomaly detection result includes: Performing anomaly detection on the third data to be analyzed using the first anomaly detection model to obtain a first anomaly detection result; and performing anomaly detection on the third data to be analyzed using the second anomaly detection model to obtain a second anomaly detection result; An abnormality detection result is obtained according to the first abnormality detection result and the second abnormality detection result.

12. The alarm method according to claim 11, characterized in that: The obtaining an abnormality detection result according to the first abnormality detection result and the second abnormality detection result includes: Performing weighted calculation on the first anomaly detection result and the second anomaly detection result to obtain an anomaly detection result; wherein the weight used in the weighted calculation is dynamically adjusted based on at least one of the following factors: data quality of the third original data; The data type of the third original data; performance evaluation results of the first anomaly detection model and the second anomaly detection model; Time passes; Working scenario of the industrial equipment; The maintenance status of the industrial equipment.

13. A computer program product, characterized in that The method comprises computer program instructions, and when the computer program instructions are read and executed by a processor, the method according to any one of claims 1 to 12 is executed.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are read and executed by a processor, the method according to any one of claims 1 to 12 is executed.

Images