Method and system for estimating product duration and performance thereof in corresponding life cycle

Abstract

The invention relates to the application of artificial intelligence and advanced statistical methods in a single software tool to monitor and analyze quality issues throughout the life cycle of a product, the whole life cycle being:-zero kilometer; -the effective service life within the warranty period ("on-site warranty period"); -an effective service life not within the warranty period ("outside warranty period"). In addition, the invention further provides an innovative method for monitoring products beyond the warranty period based on the network capture technology and artificial intelligence.

Description

TECHNICAL FIELD

[0001] The present invention relates to a method and system for estimating the duration of a product and its performance over the corresponding life cycle. BACKGROUND

[0002] As known, the life cycle of a product can be divided into three macro-phases, for example for a railway vehicle:

[0003] 1. recent production and imminent entry into service ("zero mileage");

[0004] 2. active life in warranty ("on-site warranty"); and

[0005] 3. active life out of warranty ("off-site warranty").

[0006] There are currently software tools for monitoring and analyzing zero mileage non-conformities. Software for managing claims during the manufacturing process allows the collection of information relating to the problem (e.g. component code, serial number, etc.), the problem raised by the customer and the context in which the non-conformity was found (e.g. in which type of test, during assembly, etc.). Some tools also allow the creation and management of a process for collecting the analysis carried out during the problem resolution steps. Such software can be connected to other company databases (e.g. customer list, product list) and to intelligent business tools, thus allowing the creation of graphical dashboards for monitoring specific indicators.

[0007] In addition, there are also software tools that can be used to monitor the product during the warranty period. These software tools include analysis services based on warranty-related data. These data can include: geographical data, product type data, product use data and fault description texts (claims) written by the spare parts workshop. The most advanced analysis services include: classification of claims in fault patterns, display templates, predictive analysis, reliability models.

[0008] However, the current advanced methods for analyzing warranty texts have various limitations, including those described below.

[0009] Typically, each claim is analyzed using NLP ("Natural Language Processing") techniques to extract keywords—highly informative keywords—that can be used to trace back to the component's failure mode. Consequently, each claim is associated with a set of failure modes with varying probabilities, based on state-of-the-art technology. These failure modes are predetermined and typically defined by domain experts. However, these failure modes do not correspond to the categories used by technicians during the design process (FMEA and its derivatives, where FMEA stands for "Failure Modes and Effects Analysis"). Therefore, problems discovered by quality engineers during the product's warranty period must be "translated" into FMEA categories or modes before a technical solution can be implemented.

[0010] Limon, Shah, Yadav, Om Prakash, and Nepal and Bimal (2015), in their paper "Forecasting Warranty Claim by Using Past Warranty and Field Usage Data" (pp. 2477-2484), presented at the 2014 IIE Annual Meeting and Exposition, describe a study of warranty claim failures that disrupt normal operations, leading to increased operating costs, maintenance costs, and customer dissatisfaction. The authors propose a forecasting method based on component usage to predict warranty claim trends. Field reliability parameters for the product are estimated based on warranty data and usage. To obtain field usage data, a parameter called a "product usage curve" is used, which is treated as a random variable to manage variability in product usage models. Warranty claims are forecasted for each time period using the estimated reliability parameters, the usage curve using customer reliability, and a risk function.

[0011] As for monitoring product quality issues after the warranty period ends, the inventor believes that there is currently no known method.

[0012] Finally, no single tool can monitor and analyze quality issues throughout a product's entire lifecycle.

[0013] US7516175B1 relates to a method for collecting, analyzing, and displaying warranty product-related data. It provides a graphical interface that allows for filtering and displaying data. It also allows for selecting a report model from a list of pre-installed models or creating new ones. The method does not detail the analysis performed or the type of warranty data collected. The tone is very general, likely due to the fact that the related application dates back to 2003.

[0014] US Patent No. 8,065,334 B2 describes a tool for monitoring product quality issues during the warranty period. The initial data includes data selected from the following: warranty claim data, traceability data, supplier data, manufacturer data, seller data, customer data, component data, service data, failure data, field data, fault codes, and so on. The tool itself integrates multiple functions, including the same as the present invention: natural language text analysis and the resulting structured information, reliability product engineering, and predictive analytics. However, the tool does not include analysis of newly produced parts or out-of-warranty parts, and does not detail reliability product engineering and predictive analytics.

[0015] Document CN111881259 describes a method for automatically analyzing the frequency of electrical equipment failure modes, based on text analysis of log files from running tools. This method establishes a correspondence between keywords found in text describing a failure and the failure modes described by product experts. Groups of keywords correspond to each failure mode, with each keyword's weight proportional to the probability associated with that failure mode. A "bag-of-words" model is applied to each text to convert each claim into a vector, where each dimension represents a word appearing in the claim, and its assigned value represents the frequency of that word's occurrence in the claim. The words in the claim form a multidimensional space. Each failure mode can be represented by a vector in an n-dimensional space. Domain experts fill in each dimension with a value corresponding to the probability, which is typically normalized to 1, indicating that the word corresponding to that dimension represents the failure mode in question. The text of a single claim can be represented as different vectors in the same n-dimensional space. In document CN111881259, specific text is assigned to one or more failure modes by comparing the vectors in the n-dimensional space (dot products between the vectors). US Pat. No. 7,315,861 B2 proposes a method for automatically browsing websites such as forums, blogs, and social media sites to search for users discussing products or services relevant to a business offer. This method utilizes natural language processing (NLP) technology to analyze the text of online posts and comments to identify the most profitable customers. Based on this collected data, customized sales scripts and direct marketing materials are automatically generated to attract the stated or implicit interest of the identified customers.

[0016] For this reason, there is a need for a method capable of monitoring, analyzing and managing complaints, especially those regarding mechanical component failures, based on data collected throughout the product life cycle, including from official and unofficial channels, and especially from data outside the warranty period, and therefore also from data outside the official channels of customer complaints, i.e., online channels.

[0017] In particular, it is necessary to estimate the effective duration of a product in use, as product performance can change during use. In this sense, it is also necessary to verify that the actual performance of a product or component is consistent with its theoretical design performance. Indeed, if an automotive component is designed to perform well for a certain mileage, it is necessary to estimate the time until its performance degrades due to wear, breakage, and / or failure, thereby estimating its remaining useful life.

[0018] There remains a need for a method and system for determining the duration of a product and its performance during a corresponding life cycle, which method and system are efficient and also allow for determining quality issues of the product at various life stages, in particular in the automotive field, such as products for braking systems. Summary of the Invention

[0019] The object of the present invention is to provide a method and a system for estimating the duration of a product and its performance during the corresponding life cycle, which method and system are particularly capable of monitoring, analyzing and managing data beyond the warranty period and thus beyond official channels of customer claims.

[0020] A method and a system according to the appended claims are subject matter of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The invention will now be described by way of non-limiting examples and with particular reference to the accompanying drawings, in which:

[0022] - Figure 1 shows an example of a high-level architecture of a system according to the present specification;

[0023] - Figure 2 An example of the logic flow of a method for processing a claim according to the solution of this specification is shown;

[0024] - Figure 3 shows an example of reliability model output, in particular, a time trend (dark grey line) and a predicted trend (light grey line) of the actual number of appeals received after December 2020, and corresponding confidence intervals (lines with hollow balls), according to an exemplary aspect of the invention;

[0025] - Figure 4 showing a combined severity index representation of the complaints; and

[0026] - Figure 5 An example of a computerized system for implementing the method of the present description is shown.

[0027] It should be noted that elements of different embodiments may be combined to provide further embodiments without imposing any limitation on the technical concept of the present invention, as those skilled in the art will directly and unambiguously understand from the description.

[0028] In addition, this description also refers to the prior art for implementing the present invention, and for detailed features not described, such as minor elements that are generally used in similar solutions in the prior art, reference is also made to the prior art.

[0029] When introducing an element, it is always intended to mean either "at least one" or "one or more."

[0030] When elements or features are listed in the present disclosure, they should be understood as “including these elements” or “consisting of these elements” in accordance with the present invention.

[0031] When features are listed in the same sentence or bullet list, one or more individual features may be included in the invention without being associated with other features in the list.

[0032] Two or more components (elements, devices, systems) described below may be freely combined and considered as a kit of parts of the present invention. DETAILED DESCRIPTION

[0033] First, we describe a method for automatically processing claims, which are text related to a service or product written in natural language and possibly in different languages. Claims can be generated or acquired at each of the product lifecycle steps described above.

[0034] According to the example of this specification, the overall architecture of the software system is as follows: Figure 1 As shown. Architecture 100 includes: a first module 110, which is related to managing issues with recently completed products (e.g., zero-kilometer vehicles); a second module 120, which is used to manage issues outside the warranty period; and a third module 130, which is used to manage issues within the warranty period. In managing issues within the warranty period (which can also be used in other analysis steps, such as issues outside the warranty period), there are: a module 140 for classifying claims by failure mode; a module 150 for applying predictive models; and an optional module 160 for applying reliability models. The output of one or more of these modules is passed to a display module 170.

[0035] More specifically, refer to Figure 2 , the natural language appeal 210 is in the "pre-processing" step 220, as follows:

[0036] Translation into English (or other predetermined language); translation is necessary because the libraries required for subsequent analysis need to process text in that language;

[0037] Removal of so-called “stop words”, i.e. a group of words in the text that are not important for the problem being addressed (e.g. conjunctions, prepositions, articles, etc.);

[0038] Remove punctuation marks;

[0039] Remove numbers and special characters; and

[0040] "Stemming" (keeping only the base word in a compound word).

[0041] For each word i contained in the body of the appeal processed in this way, the TF_IDF index is calculated using a state-of-the-art method that takes into account the frequency and importance of each word in the body:

[0042] tf_idf i =tf ij *idf i

[0043]

[0044] Among them, n ij is the number of times word i appears in category j, d j is the number of terms in category j, and |D| is the number of pre-defined categories for a specific product problem.

[0045] The words with a higher TF_IDF index form the so-called "keyword" set, i.e. words that are more informative about the problem being managed. The keyword list (usually not hierarchical) is regularly updated in the database in order to also take into account the latest available data.

[0046] The next step involves associating each extracted keyword with one or more FMEA categories, ie, failure mode categories that may be related to the keyword.

[0047] The association operation is performed automatically by using probability maps compiled by one or more domain experts.

[0048] A probability graph is a document (table) in which groups of failure modes are associated with each keyword extracted from the body of the claim. Each failure mode has a weight proportional to the probability of the keyword being associated with the failure mode. These weights are predetermined by domain experts based on their expertise in the field.

[0049] For each keyword, the weights associated with the failure mode add up to 1, as shown in the following example:

[0050]

[0051] Map the issues described in the body of the appeal to the failure modes present in the FMEA, allowing for:

[0052] Leverage the company’s expertise in the product and its process operations.

[0053] More accurate identification of the root cause of the indicated problem and faster selection (by operator, on a case-by-case basis) of the corrective action that needs to be implemented (this step is not included in the solution of this instruction manual);

[0054] Speed ​​up risk analysis because risk information can be directly obtained from FMEA.

[0055] Compare the number of occurrences of the same failure mode across different components / products, etc., to determine which components / products are most susceptible to a particular risk (or issue).

[0056] Analyze the impact of specific failure modes at different times in the product lifecycle using specific metrics (e.g., MERF at 3 months, 12 months, 24 months, or other regular intervals).

[0057] Automatically update the “Frequency” parameter in the FMEA analysis using the collected information.

[0058] Identification by operators of failure modes and effects that are not currently included in the FMEA.

[0059] Among the specific indicators (conventional KPIs) mentioned above, MERF (Manufacturing Engineering Repair Frequency) represents the number of claims for vehicles on each production date. Assuming i represents the production month, x represents the number of claims, y represents the number of vehicles produced and sold, and n represents the repair date minus the warranty effective date, the MERF calculation is as follows:

[0060] MERF i；3个月 =(x i；n≤3个月 / y i )·100

[0061] The value is equal to 12 months or 24 months or other predetermined interval.

[0062] Regarding the "frequency" parameter, a "frequency" is defined for each failure mode in the FMEA analysis. The frequency represents the historical number of occurrences of that failure mode (e.g., 1 occurrence per 100,000 units sold). The method of this specification allows the number of occurrences of each failure mode to be extracted from the data, thereby extracting the frequency. The frequency is then "combined" with two other parameters—gravity and detectability—to determine intervention priorities (see the AIAG & VDA FMEA manual).

[0063] Once the probability map is obtained, each claim in the corresponding failure mode can be classified 230. The specific operation is as follows:

[0064] For each keyword found in the text of the claim, the weight distribution of the corresponding failure mode is saved in the probability graph. The weights of the different failure modes are added to the keywords included in the claim. The failure mode with the higher weight is the relevant category for that claim. The table below shows an example of this process, where the claim states "The clutch was bled several times but still did not work. The slave cylinder was found to be squeaking and noisy, so it was replaced. It was then discovered that the clutch master cylinder was sucking air, and the customer also reported a fluid leak in the clutch reservoir." The keywords reported in the table below are taken from this claim:

[0065]

[0066] The fault categories are weighted as shown in the following table:

[0067]

[0068] Therefore, since the "noise" fault mode has a higher weight, this fault mode is saved in 240. Advantageously, but not limitedly, a second fault mode with a higher weight is also saved.

[0069] This is particularly beneficial in two situations:

[0070] 1. The two saved failure modes correspond to the concurrence of two failure modes in the same claim.

[0071] 2. The most likely failure mode corresponds to the “other” category (which is vague in description) and therefore does not provide any useful information in the claim category (e.g., “loss” could mean loss of fluid or loss of pressure).

[0072] Using probabilistic graphs instead of AI-based classification methods has the advantage of being able to correctly classify claims containing uncommon keywords. Given that, based on the owner's experience, most claims are categorized into a few primary failure modes for each part, a supervised AI algorithm would clearly not be able to correctly classify uncommon keywords.

[0073] Analysis of the Zero Kilometer Problem

[0074] Zero-kilometer problems, i.e., problems that occur in the manufacturer's factory or are discovered on the customer's production line and for which the manufacturer is responsible, are analyzed as follows.

[0075] The starting data involves claims that are discovered at the manufacturer's facility or by customers on their production lines, for which the manufacturer is responsible. This data may include metadata such as one or more of the following: data about the customer who filed the claim or the production line where the claim was discovered, information about the existing parts (production date, part code, etc.), the date the problem was discovered, the number of defective parts in the claim, and a preliminary description of the problem encountered in natural language, as the actual text of the claim.

[0076] After the return of the defective material, each product will undergo an in-depth analysis. In this center, the available data related to each complaint will be made available to at least one operator, who will complete the process by entering further information generated by the analysis center into a common database, including, for example, one or more of the following: a detailed description of the problem after in-depth analysis, a classification of the problems found by type and severity, the containment measures selected, a list of the analyses performed and their results, the root causes identified, the responsibility / source of the problem, the containment measures and corrective actions selected, checks to be performed on the effectiveness of the corrective actions, and standardization activities to be carried out.

[0077] For each action to be implemented, a "Task" object is created in the software, representing the specific activity to be completed. Each task has defined parameters, including at least: the owner / responsible person, the due date, and a description of the actions involved.

[0078] The method described in this manual can be applied to a preliminary description of the problem encountered written by the manufacturer / customer plant and / or a detailed description written by the internal quality operator after an in-depth analysis. The output of the analysis is a classification of the text into FMEA categories.

[0079] The starting data enriched by the method (the original text of the complaint and the text derived from the in-depth analysis) is then used to calculate specific quality indicators, such as the PPM (parts per million, e.g., ppm per part per month, etc.) of the parts for which customers have filed complaints. The collected data and calculated indicators can also be automatically displayed in a dashboard (graphical interface) through graphical displays, such as Pareto charts to identify the most complained parts, recurrence diagram problems, and the most influential complaints.

[0080] Analysis of problems during the warranty period

[0081] The analysis of quality issues during the warranty period will be carried out in accordance with the above-mentioned zero-kilometer method.

[0082] The initial data includes text requests sent by parts repair shops related to their own products under warranty. These texts are presented in natural language and different pre-set languages.

[0083] The analysis results can be displayed graphically, for example via a graphical dashboard or a web application.

[0084] In addition, such information may be linked with other data (if any), including:

[0085] Metadata of received claims stored in the database, for example including one or more of the following: customer origin of the claim, geographical area, date of visit to the repair shop / workshop and date of repair, warranty effective date, customer code of the repaired or replaced part, age of the vehicle in terms of usage or wear and tear (mileage in the case of automotive products).

[0086] Company information stored in a database includes one or more of the following: sales volume of a product over time, repair costs, product codes, and bills of materials.

[0087] By timely associating (table join operation) the above data, a particularly complete database on quality issues during the warranty period can be obtained.

[0088] The data contained therein can be displayed in various ways. Furthermore, it is possible to create indicators of interest and automatically display their values. These indicators include many classic indicators, such as one or more of the following: standard indicators, such as MERF (Manufacturing Repair Frequency): this indicator shows the number of claims for vehicles per production date; ACPU (Average Cost per Unit): the average unit cost, i.e., the total cost of warranty expenses for vehicles produced in a specific period relative to the number of vehicles produced in the same period); and time to failure.

[0089] Finally, the greater the amount of information contained in the database, the more complete the solution to quality problems. By filtering the data, it is possible to observe indicators of interest within the scope of specific customers, parts, etc.

[0090] The monitoring of quality issues during the warranty period is completed by the forecasting part, which includes:

[0091] 1. Advanced forecasting models that allow forecasting the number of future claims (of all types, such as claims for defective parts) based on data on products sold and claims over a period of time.

[0092] 2. Advanced reliability models that predict the probability of a product failing after a specific service call based on claim data over time (to predict future trends in the number of claims) and the mileage of each vehicle (generally referring to the age of the product) at the time of the warranty intervention (or claim).

[0093] The reliability model includes the following steps:

[0094] 1. Estimating the mileage of sold but unclaimed parts (censored data) based on the demand using parametric distributions;

[0095] 2. For each failure mode, insert the parameter distribution and then calculate the following statistical function (existing technology itself):

[0096] af(t) → probability density (pdf)

[0097] b.

[0098] cS(t)=P(T≥t)=1-P(T <t)=1-F(t)→生存函数(即生存函数(也称为可靠性函数)表示故障在一定时间段t后发生的概率(或等效地,部件“存活”到时刻t的概率),其中T是代表“死亡”时间(即在本情况下,故障发生的时间)的随机变量,P是概率函数,严格来说,P是被限定在测量空间上、且取值范围为[0,1]的函数;

[0099] d.

[0100]

[0101] e.

[0102]

[0103] 3. The single failure mode and its related distribution estimated in point 2 are combined by the following formula (itself a prior art):

[0104] a. Among them, S i (t) is the survival function of the ith failure mode

[0105] 4. Based on the calculation in point 3, the probability of failure as a function of mileage can be estimated.

[0106] Using the processed reliability model and the product's mission profile (claims based on a parametric distribution, estimated mileage for parts sold but not claimed (censored data)), the number of claims likely to be received in the future can be calculated using standard methods. Figure 3 An example of the reliability model output is shown in Figure 1. By anticipating demand situations, sufficient feedback on the product's service life and reliability can be collected without waiting too long, allowing for earlier intervention in the redesign of products and production chains, as well as the planning of corrective measures for products already in use.

[0107] Compared to predictive models (based on years of ownership or the age of a part or component), reliability models are more accurate and are particularly suitable for products where failures are primarily due to physical wear and tear, for example, products related to vehicle braking systems (measured per kilometer).

[0108] It should be noted that the number of claims over time is closely correlated with the number of verified failures in a component or product, and therefore provides an indication of the reliability, performance, and duration or service life of the product itself. Therefore, the solution presented in this specification addresses the problem of determining or estimating a product's performance and duration or service life beyond the scope of production and testing, and has real-world applications. This information can help improve product production technology and design to enhance its reliability and service life.

[0109] In each embodiment of this specification, "performance" may be considered to be described by or associated with metadata, demands, failure modes, future estimates of expected number of failures, etc., as described herein.

[0110] It should be noted that there may be cases where a component is replaced without a corresponding claim. However, the percentage of such unclaimed replacements is usually known or can be estimated, so that in a subsequent step or based on the method, the service life of the component can be estimated.

[0111] It should be taken into account here that "service life" can be understood as the average life, or the average life with a predetermined failure interval / type.

[0112] Analysis of out-of-warranty issues

[0113] Finally, a method is proposed for monitoring quality issues during the service life of non-warranty products (i.e., the “out-of-warranty” phase), or issues related to products that are still within the warranty period but for which repair / replacement interventions are not covered by the warranty.

[0114] Web scraping techniques can be used to explore web pages (forums, blogs, social media) that contain threaded discussions about related products. The text of the post that initiated the thread and the associated comments will be stored in the database along with metadata such as the URL, post title, comment author, and date. This specification also includes the ability to automatically analyze videos to extract text and metadata from them, known as "web scraping." The following table shows an example of a dataset extracted from a web page:

[0115]

[0116] Discussion thread articles in professional magazines or text generated from online video audio and related comments (all objects are included in the web crawl) can also be considered as discussion threads.

[0117] Formally, the text of the first element of a chain message is treated the same as a comment.

[0118] The text is processed according to the above method for use in the other product steps. For each review, the result is a list of keywords related to the failure mode and their weights, as well as the review's associated categories and their weights, stored in a database. Furthermore, for reviews obtained from web pages, it is best to extract additional information directly from the text that belongs to predefined categories relevant to the product in question. For example, in the automotive parts domain, such categories include at least: car make, vehicle model, mileage, failure descriptor, sentiment, and part name.

[0119] For each predefined category, the database provides a list of words contained in that category (e.g., the "Manufacturer" category may contain the name of a vehicle manufacturer) and / or one or more rules allowing the identification of character strings belonging to that category in the text (e.g., if the following pattern is found in the review text, i.e., a number followed by a unit of length, such as "5 km" or "12 miles", then the "Mileage" category is filled).

[0120] The output of this process is a table with a column added to the initial dataset containing the identified keywords and a column for the sentiment determined based on those keywords. For each review and each category, all the words belonging to that category and the number of times they appeared in the review are saved.

[0121] To extract information belonging to predefined categories from each threaded message text, automatic extraction algorithms can be used by searching for alphanumeric sequences or by using machine learning algorithms, specifically "Named Entity Recognition" (NER). Named Entity Recognition (NER) is an information extraction technique that attempts to identify named entities mentioned in unstructured text and classify them into predefined categories, such as names of people, organizations, places, codes, time expressions, quantities, etc. The trained algorithm can analyze each comment and extract words belonging to different predefined categories.

[0122] Then, a state-of-the-art sentiment analysis algorithm is used to calculate a sentiment confidence score for each review.

[0123] The sentiment confidence score is an objective numerical indicator of how neutral, positive, or negative a text is.

[0124] Examples (as shown in the table below):

[0125] {'positive': 0.31, 'neutral': 0.17, 'negative': 0.51}

[0126] In the claimed method, numerical calculations are first performed using a known algorithm, and the output is then refined through a series of iterations aimed at identifying specific elements associated with potential product problems.

[0127] During the prediction iterations, we take the following approach:

[0128] - Increase the negative sentiment score if the text contains specific keywords identified by the business;

[0129] - If the text consists of very few words, increase the neutral sentiment score.

[0130] We calculate a sentiment confidence score for each review / text. The sentiment level with the higher score (positive - negative - neutral) is considered the overall sentiment of the review / text.

[0131] Finally, we calculated the negative sentiment of each post. A post is a collection of comments related to the same discussion on the same website. The negative sentiment of a post is the percentage of comments with an overall negative sentiment rating compared to the total number of comments on the same post.

[0132] This emotional parameter helps determine intervention priorities.

[0133] As a result, the unstructured information contained in the text extracted from the network is structured (populated with categories) and enriched (failure mode + sentiment), and stored in the database, forming the basis for processing the monitoring indicators of out-of-warranty quality issues. The output of this processing is shown in the table above.

[0134] The data extracted from each text segment in the discussion chain message can be combined, and the combined information of each chain message can be extracted, for example:

[0135] The number of comments following the first chained message element;

[0136] The difference between the chain message start date and the last comment date;

[0137] Global sentiment obtained by averaging the sentiment of each text in the chain message;

[0138] The number of times each keyword is repeated in the chain message;

[0139] The main failure modes or sets of failure modes in descending order of probability.

[0140] The following table shows an example of the summary output (split horizontally into two parts):

[0141]

[0142] Starting from the combined data of individual chain messages, texts with negative sentiment are filtered out, and indicators for monitoring quality issues can be processed, for example, including one or more of the following parameters:

[0143] The proportion of negative chain messages to the total number;

[0144] A severity index that combines the duration of the chain message, the number of interactions, and the sentiment index (see Figure 4 Such metrics can be used to identify the most critical chain messages (i.e., those with a long duration, a high number of interactions, and very negative sentiment, thereby prioritizing them for action) or to screen different products for comparison. They can be linked to physical failures through offline FMEA analysis conducted by operators.

[0145] For each chain message, the number of interactions per unit time.

[0146] Keywords primarily associated with a specific failure mode. This metric helps understand which words are used by the same users to describe a specific problem.

[0147] The graphical dashboard can display one or more values ​​of the above indicators, including their changes over time, and can filter all available fields. In the dashboard, detailed information of all texts belonging to a specific chain message (thread) can also be displayed.

[0148] Such tools have the following advantages and allow:

[0149] Anticipating problems. Repair shop / retailer complaints are often submitted regularly (sometimes annually). Furthermore, it's not always possible to extract clear information about the problem from the complaint text. Therefore, the methods described for out-of-warranty situations complement those outlined in the key points for zero-kilometer or in-warranty situations.

[0150] Gather knowledge about out-of-warranty product performance that would otherwise be unavailable.

[0151] · Respond quickly to issues that arise to avoid negative impacts on the company and its brand image.

[0152] Monitor changes in product perception and physical state over time.

[0153] The output of the methodology for each product state (zero mileage, warranty, out-of-warranty) can be made available through a single display tool (dashboard or web application), creating a single tool for monitoring quality-related data at all stages of the product lifecycle. This tool, designed specifically for this purpose, offers the following advantages:

[0154] Comprehensive understanding of customer-perceived quality levels;

[0155] Comprehensive understanding of customer experience to improve products;

[0156] Comprehensive understanding of the quality levels of a specific component throughout its lifecycle – from production to end of warranty; and

[0157] Identify the stages in the product life cycle that are most prone to problems or physical issues leading to customer dissatisfaction so that resources can be allocated correctly and improvement plans can be developed.

[0158] Compared with the technology in the above-mentioned document CN111881259, in this specification, associating one or more failure modes with claims is not achieved through vector representation in a multidimensional space, but by extracting keywords from the claims and mapping them to failure modes, using probability graphs compiled by domain experts, thereby significantly simplifying calculations and improving estimation efficiency.

[0159] More implementation details

[0160] The method is carried out as follows.

[0161] For zero-kilometer products, data related to claims is stored in a database. All data generated by analyzing the components is collected through a graphical interface and saved in a database, where at least one key is associated with a table containing data related to the claim. The claim processing algorithm can be implemented in two modules: one for data preparation and the other for classification. Usually, the second module is functionally dependent on the results of the first module, so the two modules run in series through a data pipeline. The two modules are managed by a "coordinator" that is responsible for monitoring their life cycle and operation.

[0162] For products under warranty, claim data is stored in a database. The algorithm for processing claims can be implemented in two modules: one for data preparation and one for classification. Typically, the functionality of the second module depends on the results of the first, so the two modules run in tandem through a data pipeline. These two modules are managed by an "orchestrator" that monitors their lifecycle and operation.

[0163] The prediction algorithm can be implemented directly as a function in the display instrument. Both can be executed in parallel, that is, independently of each other.

[0164] To summarize the above, we can say that this application uses a microservices architecture. The input can come from an API or be stored in a database and / or spreadsheet. The output of the application can be provided through an API, database, spreadsheet, or dashboard.

[0165] The prediction algorithm can be encapsulated in a module and scheduled in a cluster different from the one where the application of the product under warranty is located. This is very advantageous if it is desired to ensure that the application is logically and securely isolated from the one displayed by the product at zero kilometers.

[0166] For products out of warranty, the module responsible for collecting and processing network comments can be developed in the cloud. Network data scraping can be done through scripts or by calling the API of third-party services. Network scraping is performed periodically through a cron job that fetches the comment text in raw format. The activity of importing data activates the pipeline for processing and refining the information. The output is saved in a persistent data lake where it can be viewed in detail on request.

[0167] For the calculation of the number of claims using the time function prediction model described above, in 150, the number of expected claims 335 over time for the product can be calculated by a statistical prediction algorithm based on the probability weight of each failure mode and the failure mode with the greater weight in the respective text of one or more previously managed claims.

[0168] To implement the calculation 160 of the probability of failure over time for the product, for a predetermined failure mode, it can be calculated based on the number of claims 335 over time and the product use time extracted from the metadata of said one or more claims. A possible mathematical calculation model is known and described in the article by Limon et al. (section 2.4) above, but obviously only applicable to this calculation since, as we have explained, the overall method is problematic. Other implementation techniques are feasible according to the current and future state of the art.

[0169] This type of calculation 160 can be repeated for each failure mode and then the respective probabilities of failure can be linearly combined, obtaining the overall probability of failure over time for the product.

[0170] All the implementation methods described above can be stored in a local server and in the cloud, also in a hybrid mode.

[0171] Reference Figure 5 is shown a possible hardware implementation of the method of the application. The system 300 comprises a server 320, a database 330 connected to said server and a client 310 connected to said server through a telecommunications network 340, conveniently through the Internet.

[0172] The server 320 is configured to perform the network scraping according to the description above, while the client 310 is equipped with a graphical interface 315 to perform the display steps described above.

[0173] The database 330 is configured to store the output of one or more steps of the above method, in particular the various claims 335 (also obtained from web scraping). The database can be contained in the server or can be remotely connected to the server. In addition, no direct connection is specified between the client 310 and the database 330, but a connection between the client 310 and the database 330 can be conveniently provided.

[0174] The system described above allows monitoring of components for signs of problems from the initial steps of production until the end of the warranty period. It can also be extended to steps prior to production and includes software that allows management of critical issues that arise during component design and validation.

[0175] An innovation of the method is the ability to utilize Figure 2 The same approach / algorithm described in [1] can be used to analyze issues raised at different lifecycle stages and from completely different data sources. Using the same approach of identifying specific failure modes based on keywords in the text, it is possible to analyze a product's performance across the entire lifecycle with respect to specific issues and determine at which stage the primary critical issues arise. The same algorithm can also be used to simultaneously study different stages of a product and return an overall combined output of component status for analysis.

[0176] It is not common practice to apply the same algorithm / method to different stages of the life cycle. The demands involving different stages of the product life cycle have the following characteristics:

[0177] - Different data collection methods

[0178] -Different data sources

[0179] - Different structures

[0180] - Different data reliability

[0181] - Different data integrity

[0182] The method is able to adapt to and manage the peculiarities of each different data source and structure. The method is able to automatically process different data sources and create a structure that allows them to be linked together using the same algorithm that is appropriately designed for raw data of diverse structure and content.

[0183] The above describes the preferred embodiments of the present invention and proposes some variations of the present invention, but it should be understood by those skilled in the art that modifications and changes can be made without departing from the relevant protection scope defined by the appended claims.

Claims

1. A method implemented by a computer (310, 320, 330, 340) for estimating the duration of a product and the performance of the product during its corresponding life cycle, the method comprising the following steps: A. Obtaining one or more user claims for the product (335) and storing the one or more claims in a database (330), wherein the one or more claims each include corresponding text and corresponding metadata, and obtaining the usage time of the product from the corresponding text and the corresponding metadata; B. analyzing the corresponding texts ( 110 , 120 , 130 ), and establishing a corresponding keyword list for each corresponding text; C. for each keyword in the corresponding keyword list in step B, assigning (140) a corresponding probability weight to each failure mode in the predefined failure mode list, and storing the assigned probability weight in the database (330); D. For each failure mode, summing the probability weights assigned to the failure mode for each keyword in step C for each corresponding text (140); E. Selecting the failure mode with a larger weight obtained in step D for each corresponding text and saving the selected failure mode in the database (330); F. using a statistical prediction algorithm to calculate (150) the number of expected claims (335) for the product over time based on the probability weight of each failure mode and the probability weight of the failure mode with a larger weight according to the corresponding text of the one or more claims in steps A to E; G. Calculating the useful life of the product based on a reliability model, the reliability model taking the expected number of claims (335) and usage time of the product in step F as input; wherein the one or more appeals (335) include at least one appeal consisting of text obtained by web crawling (320) and corresponding metadata extracted from the text; and wherein the following additional steps are performed: H. Displaying one or more pieces of information calculated in steps B to G on a graphic interface (170, 315).

2. The method according to claim 1, wherein In step F, the statistical algorithm uses the formula To calculate the expected number of appeals, where S i (t) is a survival function representing the probability that a failure of the i-th failure mode will occur after a time period t, and the survival function is expressed as follows: S(t)=P(T≥t)=1-P(T<t)=1-F(t) Wherein, T is a random variable representing the time when the failure of the i-th failure mode occurs, P is a probability function whose value is in the range of [0,1], and F(t) is a distribution function expressed as follows: where f(t) is the probability density associated with the random variable T.

3. The method according to claim 1 or 2, wherein: The product is a component of a motor vehicle.

4. The method according to claim 3, wherein: The component of the motor vehicle is a component of a brake system.

5. The method according to claim 1 , wherein The metadata includes one or more of a product manufacturer, a product model, a registered product lifespan, a product name, and a product mission profile.

6. The method according to claim 1, wherein The failure mode is a failure mode and effects analysis (FMEA) mode, and the failure mode and effects analysis mode includes FMEA parameters related to failure causes and solutions.

7. The method according to claim 1, wherein In step E, a second failure mode with a second larger weight is additionally selected and stored, and steps F and G are also performed for the second failure mode, or, if the first failure mode is a normal mode in the failure mode list, steps F and G are performed only for the second failure mode.

8. The method according to claim 1, wherein After manual inspection, an analysis text with relevant metadata for the product is obtained, and the analysis text with relevant metadata is processed into a claim according to steps A to E, and then steps F and G are performed again for each corresponding text according to steps A to E and for the analysis text based on the keyword list, the corresponding probability weight, and the failure mode with a larger weight.

9. The method according to claim 1, wherein In step H, one or more of the corresponding metadata are additionally displayed.

10. The method according to claim 1, wherein: Relevant metadata is extracted from texts obtained by web scraping by using machine learning algorithms, in particular "Named Entity Recognition" (NER).

11. The method according to claim 1, wherein: For each corresponding text, a numerical indicator indicating an evaluation of the product is determined by an algorithm, the numerical indicator is stored in the database, and the numerical indicator is used in step H, wherein the algorithm is configured to calculate the numerical indicator of sentiment analysis based on the score of the words of the corresponding text.

12. The method according to claim 11, wherein Prior to step F, the severity index of the fault is calculated based on parameters indicating user emotions, and in the case of text obtained by web crawling of chain messages, the severity index of the fault is also calculated based on the duration of the chain messages, the number of interactions and the fault mode.

13. The method according to claim 12, wherein: Repeat step H for each failure mode, and then linearly combine the corresponding failure probabilities to obtain the overall failure probability of the product as it changes over time.

14. A system (300) for determining the duration of a product and the performance of the product during a corresponding life cycle, the system (300) comprising: - server (320); - a database (330) connected to said server; - a client (310) connected to said server via a telecommunication network (340); in: - the server (320) is configured to perform web crawling and steps A to G according to one of claims 1 to 13; - the client (310) is equipped with a graphical interface (315) to perform step H according to one of claims 1 to 13; - the database (330) being configured to store the outputs of steps A to G according to one of claims 1 to 13.

15. The system according to claim 14, wherein: The analysis text according to claim 8 is input into the client (310) and transmitted to the server (320) via the client.

Images