Anomaly prediction system, anomaly prediction method, and anomaly prediction program

Abstract

During model operation, it will be possible to control how far in advance the inference results are sent to the maintenance site from the day the target event is expected to occur. [Solution] The anomaly occurrence prediction system 1 includes a provisional model creation unit 207 that creates a provisional model for calculating the probability of a target anomaly occurring using a training dataset, a group creation unit 208 that inputs verification data into the provisional model, groups the verification data based on the output results, changes the parameter M for each group, and creates a new dataset for retraining, and an operational model creation unit 209 that formally creates an operational model 104 for calculating the probability of a target anomaly occurring using the retraining dataset.

Description

Technical Field

[0001] The present invention relates to an abnormality occurrence prediction system, an abnormality occurrence prediction method, and an abnormality occurrence prediction program.

Background Art

[0002] Patent Document 1 discloses a technique for performing maintenance before an abnormality occurs by collecting data from market-operating machines, generating an abnormality occurrence model, and using the model inference result.

Summary of the Invention

Problems to be Solved by the Invention

[0003] In the conventional method described in Patent Document 1, maintenance is performed before an abnormality occurs by collecting data from market-operating machines, generating an abnormality occurrence model, and using the model inference result, but it does not consider how many days before the day when the target event will occur during model operation the inference result should be reported to the maintenance site.

[0004] An object of the present invention is to make it possible to control how many days before the day when the target event will occur during model operation the inference result is reported to the maintenance site.

Means for Solving the Problems

[0005] To solve the above-mentioned problems, an abnormal occurrence prediction system according to one aspect of the present invention is an abnormal occurrence prediction system that predicts abnormal phenomena occurring in an image forming apparatus in advance, comprising: a market-operating machine data storage unit that collects data from within the image forming apparatus at predetermined intervals and stores it for a certain period of time; a maintenance history data storage unit that stores the maintenance history of each image forming apparatus; a learning dataset creation unit that combines data that is considered to correspond to an abnormal situation based on the date of occurrence of the target abnormal phenomenon from the maintenance history with data randomly extracted from the maintenance history of image forming apparatuses where the target abnormal phenomenon has not occurred, and generates a learning dataset based on set parameters; a provisional model creation unit that creates a provisional model that calculates the probability of occurrence of the target abnormal phenomenon using the learning dataset; and inputting a continuous dataset of the past certain period of the image forming apparatus in which the target abnormal phenomenon occurred from the maintenance history data into the provisional model. 、 The probability of the target abnormal phenomenon occurring each day is calculated, and the day on which the probability of the abnormal phenomenon first exceeds a predetermined threshold and the difference between the next few days of abnormal phenomenon occurrence are recorded. 、 The system includes a group creation unit that calculates the predicted number of days slack for occurrence from the results, groups the dataset according to the predicted number of days slack for occurrence, modifies the parameters for each group, and creates a new dataset for retraining, and an operation model creation unit that formally creates an operation model for calculating the probability of the target abnormal phenomenon occurring using the retraining dataset. [Effects of the Invention]

[0006] During model operation, it is possible to control how far in advance the inference results are sent to the maintenance site from the day the target event is expected to occur. [Brief explanation of the drawing]

[0007] [Figure 1] Hardware configuration diagram for model creation in the anomaly prediction system according to the embodiment. [Figure 2] Functional block diagram of the model training server [Figure 3] Diagram explaining how to determine the anomaly detection threshold. [Figure 4]A diagram showing an example of grouping of verification data. [Figure 5] Hardware configuration diagram of the model training server [Figure 6] Flowchart for creating an operational model [Figure 7] Configuration diagram showing an example of equipment to which the abnormal occurrence prediction system according to the embodiment is applied. [Figure 8] Block diagram showing the hardware configuration of an example of the equipment shown in Figure 7. [Modes for carrying out the invention]

[0008] The embodiments will be described below with reference to the attached drawings. To facilitate understanding of the explanation, the same reference numerals are used for identical components in each drawing whenever possible, and redundant explanations are omitted.

[0009] Figure 1 is a hardware configuration diagram for model creation in the anomaly occurrence prediction system 1 according to this embodiment. The anomaly occurrence prediction system 1 according to this embodiment is a system that predicts abnormal phenomena occurring in an image forming apparatus in advance using the created model (operation model).

[0010] As shown in Figure 1, the anomaly prediction system 1 includes, as components related to model learning, an image forming apparatus 101-1, a maintenance history device 101-2, a data storage server 102, a model learning server 103, and an operation model 104.

[0011] Image forming apparatus 101-1 is one or more image forming apparatuses operating in the market. In the example in Figure 1, image forming apparatus 101-1 includes n image forming apparatuses 1 to n. Each image forming apparatus 101-1 periodically (for example, once a day) transmits data related to equipment information to the data storage server 102. Alternatively, the data storage server 102 may be the main operator and periodically acquire equipment information from each of the image forming apparatuses 101-1. Image forming apparatus 101-1 may include, for example, printers and multifunction devices.

[0012] The maintenance history device 101-2 is a device that stores past maintenance history information for each image forming apparatus 101-1. Maintenance history information includes data such as the history of failures and abnormalities and the maintenance performed at that time. The maintenance history device 101-2 periodically (for example, once a day) transmits data related to the maintenance history information to the data storage server 102. Alternatively, the data storage server 102 may be the main operator and periodically acquire the maintenance history information from the maintenance history device 101-2.

[0013] The data storage server 102 is a device that stores and saves equipment information acquired from each image forming apparatus 101-1 and maintenance history information acquired from the maintenance history device 101-2.

[0014] The model training server 103 performs training to create a failure prediction model (operational model). The model training server 103 acquires only the data necessary for training from the data storage server 102. The data used for training could, for example, be approximately one year's worth of data, taking seasonal factors into consideration.

[0015] The operational model 104 is a model generated as a result of training the model training server 103. The operational model 104 includes a supervised learning model, such as a multi-layer neural network, and has a structure that outputs output information corresponding to arbitrary input information.

[0016] In the model training server 103, the training method used to generate the operational model 104 can be, for example, a well-known method such as backpropagation or deep learning, if the operational model 104 is a multi-layer neural network.

[0017] On the other hand, after the model is created as shown in Figure 1, the anomaly prediction system 1 extracts the necessary data from the data storage server 102 and inputs it into the generated operational model 104. This allows the system to predict in advance the probability of an anomaly occurring in each image forming apparatus 101-1 based on the output information of the operational model 104.

[0018] Figure 2 is a functional block diagram of the model learning server 103. As shown in Figure 2, the model learning server 103 includes a temporary model creation block 201, a group creation unit 208, and an operation model creation unit 209, and performs control in this order.

[0019] The temporary model creation block 201 includes a market-operating machine data storage unit 202, a maintenance history data storage unit 203, a parameter setting unit 204, a data processing unit 205, a learning data set creation unit 206, and a temporary model creation unit 207.

[0020] The market-operating machine data storage unit 202 stores various information such as the above-described device information regarding the machines (any one of the image forming apparatuses 1 to n) that are in market operation among the image forming apparatuses 101-1.

[0021] The maintenance history data storage unit 203 stores various information such as maintenance history information regarding the machines in market operation acquired from the maintenance history device 101-2.

[0022] The parameter setting unit 204 sets a predetermined parameter M. Although the details of the parameter M will be described later, generally, it includes a value related to adjustment of how far in advance to predict and issue a warning from the occurrence of an abnormality.

[0023] The data processing unit 205 processes various data including the device information stored in the market-operating machine data storage unit 202 and the maintenance history information stored in the parameter setting unit 204 for model learning. The data processing unit 205 performs data processing using the parameter M set by the parameter setting unit 204. For example, the data processing unit 205 adjusts and extracts the length of the time-series data of the device information and the maintenance history information according to the magnitude of the parameter M.

[0024] The training dataset creation unit 206 creates a training dataset for training a provisional model using the data set processed by the data processing unit 205. The training dataset created here contains a mixture of data representing abnormal occurrences and normal data. The data set processed by the data processing unit 205 is the input information for the provisional model. The output information for the provisional model corresponds to the probability of an abnormal occurrence. In the training dataset, for example, if the range of the output information is from 0% to 100% of the probability, the output information (teacher signal) corresponding to the input information in an abnormal occurrence state is set to, for example, "100". On the other hand, the output information (teacher signal) corresponding to the input information in normal data without abnormal occurrences is set to, for example, "0".

[0025] The provisional model creation unit 207 uses the training dataset created by the training dataset creation unit 206 to perform machine learning on the input-output relationships of the provisional model.

[0026] The following methods are possible for creating the model in the provisional model creation section 207. The method will be determined after considering the advantages and disadvantages of each.

[0027] • Classification-based machine learning method (gradient boosting) Advantages: Low computational load for inference during operation / Can calculate predicted values ​​with high accuracy. Disadvantage: Because the model's contents cannot be visualized, it is difficult to analyze cases where the accuracy is low.

[0028] • Classification-based machine learning methods (logistic regression) Advantages: Model contents can be visualized / Inference computation load during operation is not very high Disadvantage: The model's prediction accuracy may not be high.

[0029] • Classification-type deep learning means Advantages: It is possible to calculate predicted values ​​with high accuracy. Disadvantages: The model's contents cannot be visualized, making analysis difficult when accuracy is low / The computational load on inference during operation is high.

[0030] The group creation unit 208 inputs verification data into the provisional model created by the provisional model creation unit 207 and groups the data according to the buffer days based on the output results. Here, the buffer days (also called "predicted occurrence buffer days") are the number of days until the day when notification of an abnormality is given before the predicted date, in the case of any image forming apparatus 101-1 where an abnormality is predicted to occur in advance. The buffer days are determined appropriately for each image forming apparatus, for example, according to the status of preparations for countermeasures against abnormalities.

[0031] The "verification data" used by the group creation unit 208 for grouping is, for example, a data set of input information for a provisional model that has been processed in the same way as the training dataset by the data processing unit 205, and is different from the training dataset. The verification data also includes a dataset of continuous equipment information for a certain period in the past of the image forming apparatus 101-1 in which the predicted abnormal phenomenon occurred, as recorded in the maintenance history information. Alternatively, the data set in which abnormal phenomena occurred may be used as verification data by utilizing the input information of the training dataset used for machine learning of the provisional model.

[0032] The operational model creation unit 209 distributes the validation data into multiple groups based on the grouping information created by the group creation unit 208, and creates multiple retraining datasets for each group. Then, it further performs machine learning on a provisional model for each of the multiple retraining datasets to create multiple operational models 104. Figure 2 illustrates an example configuration in which Z operational models 1 to Z are created. The operational model creation unit 209 can create operational models 104 using the same machine learning methods as the provisional model creation unit 207.

[0033] Now, referring to Figures 3 and 4, we will further explain the grouping according to the aforementioned number of days of leeway.

[0034] Figure 3 illustrates the method for determining the anomaly detection threshold. The vertical axis of Figure 3 represents the probability that some kind of anomaly may occur in any image forming apparatus 101-1 to be judged (target anomaly occurrence probability), ranging from 0 (%) to 100 (%). The target anomaly occurrence probability is the output information when verification data is input into the provisional model. The horizontal axis of Figure 3 represents the year, month, and day. The year, month, and day information is recorded and linked to each piece of information in the verification data.

[0035] At market maintenance sites, even if an abnormal condition is predicted several days in advance, it may not be possible to go there immediately. Therefore, there is a desire to receive abnormal condition notifications with sufficient lead time (the lead time mentioned above). However, the process leading to the same abnormal condition can differ. Figure 3 shows the time-series information of the output of the provisional model, i.e., the probability of occurrence of the target abnormal phenomenon when verification data related to machines (image forming apparatus) that have previously experienced abnormal phenomena is input into the provisional model created in the provisional model creation unit 207 in Figure 2.

[0036] Figure 3 shows the time-series graphs G1 and G2 of the provisional model output for two machines, illustrated as solid and dotted lines, respectively. As shown in Figure 3, the provisional model outputs G1 and G2 for the two machines show different time-series trends, but in both cases, the probability of an anomaly occurring gradually increases from 0% to around 100%. As shown in Figure 3, the time-series graphs G1 and G2 of the provisional model output for the two machines both show the anomaly occurring on the same day (labeled "SC occurrence date" in Figure 3). Note that "SC occurrence date" can also be expressed as the date on which the predicted anomaly occurred (target anomaly occurrence date), and includes not only the day the anomaly occurred but also the day on which countermeasures were taken after it occurred (countermeasure date). On the SC occurrence date, both time-series graphs G1 and G2 show the maximum probability of an anomaly occurring, which is around 100%.

[0037] Here, as shown in Figure 3, the state deterioration threshold A (%) is set to a probability lower than the probability of an abnormal phenomenon occurring on the day of the SC occurrence, at a point before the SC occurrence date. The time series graph G1 of the hypothesis model output for one machine shows a shape in which the probability of detecting an abnormal phenomenon gradually increases over a long period of time, reaching the state deterioration threshold A (also called the "probability threshold for occurrence of abnormal phenomenon") X days before the abnormality occurs (X days before the SC occurrence date). In contrast, the time series graph G2 of the hypothesis model output for the other machine shows a rapid increase in the probability of detecting an abnormal phenomenon in a short period of time, so it reaches the state deterioration threshold A Y days before the abnormality occurs, which is shorter than X days (X >> Y). Thus, even for the same abnormal phenomenon, if the process leading to it is different, the time it takes to reach the state deterioration threshold A will be different even when using a common hypothesis model. Here, the period from the day when the probability of detecting an abnormal phenomenon (hypothesis model output) reaches the state deterioration threshold A to the day of the SC occurrence is referred to as the "margin of error" mentioned above.

[0038] Figure 4 shows an example of group division of validation data. In the example in Figure 4, the group creation unit 208 divides the time series data of each provisional model output of the validation data into one of three groups: "0-19 days", "20-40 days", and "41 days or more", according to the number of days of leeway.

[0039] In actual operation, the aim is to ensure a buffer period of days before the occurrence of an anomaly, so that the model predicts and issues a warning.

[0040] In this embodiment, the provisional model creation unit 207 shown in Figure 2 creates a provisional model for the same abnormal phenomenon all at once. Then, the group creation unit 208 shown in Figure 2 inputs verification data into the created provisional model to obtain a provisional model output corresponding to the abnormal phenomenon detection probability. From the results, it is confirmed how many days before the occurrence date of the abnormal phenomenon a prediction and warning could have been issued using the provisional model, and the data is grouped by buffer day. However, large variations may occur in the buffer days.

[0041] Therefore, in this embodiment, as shown in Figure 4, the group creation unit 208 shown in Figure 2 groups the verification data by the number of days of leeway. Then, the operation model creation unit 209 shown in Figure 2 uses the verification data set contained in each group to perform machine learning on a provisional model again for each group to create the final operation model 104.

[0042] In this process, as illustrated in Figure 4, the parameter M, explained with reference to Figure 2, is changed for each group to create the operational model 104. If the slack period was short in the provisional model, the parameter M value is adjusted in the remodel to lengthen the slack period. Conversely, if the slack period was long in the provisional model, the parameter M value is adjusted in the remodel to shorten the slack period.

[0043] As a result, the same abnormal phenomenon will be handled by multiple operational models 104. If an abnormality is triggered for each operational model 104, multiple notifications will be sent for the same abnormal phenomenon, which could result in an operational format that is difficult to use in maintenance sites. Therefore, the following countermeasures are being considered.

[0044] • If multiple operational models 104 are created for the same abnormal phenomenon, the system will trigger an alarm if even one of them exceeds the condition deterioration threshold A.

[0045] • If multiple operational models 104 are created for the same abnormal phenomenon, the system determines whether an abnormality has occurred based on the model result that outputs the highest probability among the results of the multiple operational models 104.

[0046] • If multiple operational models 104 are created for the same abnormal phenomenon, the system will trigger an abnormality alert (majority vote) if the number of models exceeding the state deterioration threshold A is greater than the number of models not exceeding the state deterioration threshold A.

[0047] Figure 5 is a hardware configuration diagram of the model training server 103. As shown in Figure 5, the model training server 103 can be physically configured as a computer system including a processor 11 that includes a CPU (Central Processing Unit) and GPU, main memory RAM (Random Access Memory) 12 and ROM (Read Only Memory) 13, input devices such as a keyboard and mouse 14, output devices such as a display 15, a communication module 16 which is a data transmission and reception device such as a network card, and auxiliary storage devices such as a hard disk 17. Each function of the model training server 103, as explained with reference to Figures 2 to 4, is realized by loading predetermined computer software (anomaly prediction program) onto the hardware such as the processor 11 and RAM 12, thereby operating the communication module 16, input devices 14 and output devices 15 under the control of the processor 11, and reading and writing data to RAM 12 and auxiliary storage devices 17.

[0048] Figure 6 is a flowchart of the procedure for creating operational model 104 (method for predicting anomalies).

[0049] In step S11, the market-operating machine data storage unit 202 and the maintenance history data storage unit 203 collect data to be used for the model phenomenon.

[0050] In step S12, the parameter setting unit 204 sets the parameter M related to the margin of days remaining since the occurrence of the abnormal phenomenon.

[0051] In step S13, the provisional model creation unit 207 creates a provisional model using the training dataset created by the training dataset creation unit 206.

[0052] Prior to step S13, the data processing unit 205 processes various data, including equipment information stored in the market-operating machine data storage unit 202 and maintenance history information stored in the parameter setting unit 204, for model training. The data processing unit 205 performs data processing using the parameter M set by the parameter setting unit 204. One example of data processing is a method of adjusting the length of time-series data of equipment information and maintenance history information according to the magnitude of parameter M. Furthermore, the training dataset creation unit 206 creates a training dataset for use in training a provisional model using the data set processed by the data processing unit 205.

[0053] In step S14, the group creation unit 208 inputs validation data (training data is also acceptable) into the provisional model created in step S13, and calculates the day on which the state deterioration threshold A is reached and the number of days until the subsequent occurrence of an abnormal phenomenon (save-day margin) (save-day calculation).

[0054] In step S15, the group creation unit 208 performs multiple groupings based on the range of slack days calculated in step S14.

[0055] In step S16, the group creation unit 208 individually sets the parameter M related to the margin of days before the occurrence of an abnormal phenomenon for each group created in step S15.

[0056] In step S17, the group creation unit 208 creates a retraining dataset from the validation data (or training dataset) for each group created in step S15, according to the parameter M set in step S16, and the operational model creation unit 209 uses the created retraining dataset to perform machine learning and create a model (operational model 104).

[0057] In step S18, the anomaly prediction system 1 combines the operational models 104 created by each group and configures the system to issue an alert if any of the conditions for an alert are met.

[0058] As described above, the abnormal occurrence prediction system 1 according to this embodiment is an abnormal occurrence prediction system that predicts abnormal phenomena occurring in the image forming apparatus 101-1 in advance, and includes a market operation machine data storage unit 202 that collects data from within the image forming apparatus 101-1 at predetermined intervals and stores it for a certain period of time, a maintenance history data storage unit 203 that stores the maintenance history of each image forming apparatus 101-1, a learning dataset creation unit 206 that combines data that is considered to correspond to an abnormal situation based on the date of occurrence of the target abnormal phenomenon (SC occurrence date) from the maintenance history with data randomly extracted from the maintenance history of image forming apparatus 101-1 where the target abnormal phenomenon has not occurred, and generates a learning dataset based on a set parameter M, and using the learning dataset The system includes: a provisional model creation unit 207 that creates a provisional model for calculating the probability of the target abnormal phenomenon occurring; a group creation unit 208 that inputs a continuous dataset (verification data) of the past certain period of time for an image forming apparatus in which the target abnormal phenomenon occurred, based on maintenance history data, into the provisional model, calculates the probability of the target abnormal phenomenon occurring for each day, records the day on which the probability of the abnormal phenomenon occurring for the first time exceeds a predetermined abnormal phenomenon occurrence threshold A and the nearest abnormality occurrence day thereafter, calculates the predicted number of days of occurrence leeway from the results, groups the dataset according to the predicted number of days of occurrence leeway, changes the parameter M for each group, and creates a new dataset for retraining; and an operation model creation unit 209 that formally creates an operation model 104 for calculating the probability of the target abnormal phenomenon occurring using the retraining dataset.

[0059] This configuration allows for the creation of a separate operational model 104 for each group by grouping verification data according to the predicted number of days before the event is expected to occur, using the provisional model created once. This enables the control of how far in advance the inference results are sent to the maintenance site during model operation, by utilizing multiple operational models 104 created in this way. Regarding the method of utilizing multiple operational models 104, various methods can be selected, as explained with reference to Figure 4.

[0060] Figure 7 is a configuration diagram showing an example of equipment to which the abnormal occurrence prediction system 1 according to this embodiment is applied. In Figure 7, an image forming apparatus 300 (corresponding to the image forming apparatus 101-1 in Figure 1) is shown as an example of equipment to which this embodiment is applied. The image forming apparatus 300 can be applied to a multifunction device having at least two functions from among a copy function, a printer function, a scanner function, and a facsimile function. Furthermore, the equipment is not limited to an image forming apparatus; other equipment besides image forming apparatuses include, for example, robots and amusement equipment. The image forming apparatus 300 shown in Figure 7 is a digital copier having a paper feeding unit 303 and an image forming apparatus body 304, with an image reading device 100 and an automatic document feeding device 200 mounted on top of these.

[0061] The main body 304 of the image forming apparatus is equipped with a tandem-type image forming unit 305, a registration roller 308 that transports recording paper supplied to the image forming unit 305 from the paper feeding unit 303 via a transport path 307, an optical writing device 309, a fixing transport unit 310, and a double-sided tray 311.

[0062] The imaging unit 305 has four photoreceptor drums 312 arranged side by side, corresponding to four toner colors: yellow (Y), magenta (M), cyan (C), and black (K). Imaging elements, including a charger, developer 306, transfer unit, cleaner, and static eliminator, are arranged around each photoreceptor drum 312.

[0063] Furthermore, an intermediate transfer belt 313 is positioned between the transfer unit and the photoreceptor drum 312, stretched between the drive roller and the driven roller while being held between the nip of both.

[0064] In the tandem-type image forming apparatus 300 configured in this way, light is written to each color, Y, M, C, and K, on ​​the corresponding photosensitive drum 312, and the toner stored in each color cartridge 314 is developed in the developer unit 306, and primary transfer is performed on the intermediate transfer belt 313 in the order of Y, M, C, K, for example.

[0065] The image forming apparatus 300 then transfers the four-color superimposed full-color image onto the recording paper in a secondary transfer, fixes it, and ejects the paper to form a full-color image on the recording paper. The image forming apparatus 300 also forms the image read by the image reading device 100 onto the recording paper.

[0066] Figure 8 is a block diagram showing the hardware configuration of an example of the equipment shown in Figure 7. As shown in Figure 8, the image forming apparatus 300, an example of the equipment shown in Figure 7, has a configuration in which the controller 410 and the engine unit 460 are connected by a PCI (Peripheral Component Interface) bus. The controller 410 is a controller that controls the overall operation of the image forming apparatus 300, as well as drawing, communication, and input from an operation unit (not shown). The engine unit 460 is an engine that can be connected to the PCI bus, such as a plotter or other print engine. In addition to the engine part, the engine unit 460 includes an image processing part such as error diffusion and gamma conversion.

[0067] The controller 410 comprises a CPU (Central Processing Unit) 411, a northbridge (NB) 413, system memory (MEM-P) 412, a southbridge (SB) 414, local memory (MEM-C) 417, an ASIC (Application Specific Integrated Circuit) 416, and an HDD (Hard Disk Drive) 418, with the northbridge (NB) 413 and the ASIC 416 connected by an AGP (Accelerated Graphics Port) bus 415. Furthermore, the MEM-P 412 includes ROM (Read Only Memory) 412a and RAM (Random Access Memory) 412b.

[0068] The CPU 411 controls the entire image forming apparatus 300 and has a chipset consisting of NB413, MEM-P412, and SB414, through which it is connected to other devices.

[0069] The NB413 is a bridge for connecting the CPU411 with the MEM-P412, SB414, and AGP415, and includes a memory controller that controls reading and writing to the MEM-P412, as well as a PCI master and an AGP target.

[0070] MEM-P412 is a system memory used for storing programs and data, expanding programs and data, and drawing data for printers. It consists of a ROM412a and a RAM412b. ROM412a is a read-only memory used for storing programs and data, while RAM412b is a writable and readable memory used for expanding programs and data, drawing data for printers, and other purposes.

[0071] The SB414 is a bridge for connecting the NB413 to PCI devices and peripheral devices. The SB414 is connected to the NB413 via the PCI bus, which also connects to the network interface (I / F) section, among other things.

[0072] The ASIC416 is an integrated circuit (IC) for image processing applications that has hardware elements for image processing, and acts as a bridge connecting the AGP415, PCI bus, HDD418, and MEM-C417. The ASIC416 consists of a PCI target and an AGP master, an arbiter (ARB) which is the core of the ASIC416, a memory controller that controls the MEM-C417, multiple DMACs (Direct Memory Access Controllers) that perform image data rotation etc. using hardware logic, and a PCI unit that performs data transfer via the PCI bus with the engine unit 460. The ASIC416 is connected to an FCU (Facsimile Control Unit) 430, a USB (Universal Serial Bus) 440, and an IEEE1394 (the Institute of Electrical and Electronics Engineers 1394) interface 450 via the PCI bus. The operation display unit 420 is directly connected to the ASIC416.

[0073] MEM-C417 is local memory used as a copy image buffer and code buffer, while HDD418 is storage for accumulating image data, programs, font data, and forms.

[0074] AGP415 is a bus interface for graphics accelerator cards proposed to accelerate graphics processing, and it speeds up graphics accelerator cards by providing high-throughput direct access to MEM-P412.

[0075] The embodiments have been described above with reference to specific examples. However, this disclosure is not limited to these specific examples. Modifications made to these specific examples by those skilled in the art are also included within the scope of this disclosure, as long as they retain the features of this disclosure. The elements, their arrangement, conditions, shapes, etc., of each of the aforementioned specific examples are not limited to those illustrated and can be modified as appropriate. The elements of each of the aforementioned specific examples can be combined in different ways as appropriate, as long as no technical inconsistencies arise. [Explanation of Symbols]

[0076] 1. Anomaly Occurrence Prediction System 101-1, 300 Image forming apparatus 202 Machine data storage unit currently in operation in the market 203 Maintenance history data storage unit 206 Training Dataset Creation Section 207 Provisional Model Creation Department 208 Group Creation Department 209 Operational Model Creation Department [Prior art documents] [Patent Documents]

[0077] [Patent Document 1] Japanese Patent Publication No. 2019-159730

Claims

1. An abnormal occurrence prediction system that predicts abnormal phenomena occurring in an image forming apparatus in advance, A market-operating machine data storage unit collects data from within the image forming apparatus at predetermined intervals and stores it for a certain period of time. A maintenance history data storage unit that stores the maintenance history of each of the aforementioned image forming apparatuses, A training dataset creation unit generates a training dataset based on set parameters by concatenating data that is considered equivalent to an abnormal situation based on the date the target abnormal phenomenon occurred from the maintenance history with data randomly extracted from the maintenance history of image forming apparatuses where the target abnormal phenomenon did not occur, and by linking this data with data that is considered equivalent to an abnormal situation based on the date the target abnormal phenomenon occurred from the maintenance history. A provisional model creation unit creates a provisional model that calculates the probability of the target abnormal phenomenon occurring using the aforementioned training dataset, A group creation unit inputs a continuous dataset of the image forming apparatus over a certain past period in the maintenance history data where the target abnormal phenomenon occurred into the provisional model, calculates the probability of the target abnormal phenomenon occurring each day, records the day on which the probability first exceeds a predetermined threshold for abnormal phenomenon occurrence and the nearest day of occurrence thereafter, calculates the predicted number of days of occurrence leeway from the results, groups the dataset according to the predicted number of days of occurrence leeway, modifies the parameters for each group, and creates a new dataset for retraining. An operational model creation unit formally creates an operational model for calculating the probability of the target abnormal phenomenon occurring using the aforementioned retraining dataset, An anomaly prediction system equipped with the following features.

2. The aforementioned operational model creation unit creates multiple operational models for each group. If any one of the aforementioned operational models exceeds the condition deterioration threshold for the same abnormal phenomenon, an abnormality alert will be triggered. An anomaly prediction system according to claim 1.

3. The aforementioned operational model creation unit creates multiple operational models for each group. For the same abnormal phenomenon, the abnormality is detected based on the model result that outputs the highest probability among the results of the multiple operational models mentioned above. An anomaly prediction system according to claim 1.

4. The aforementioned operational model creation unit creates multiple operational models for each group. An abnormality alert is triggered when, for the same abnormal phenomenon, the number of models in the aforementioned operational models that exceed the condition deterioration threshold is greater than the number of models that do not exceed it. An anomaly prediction system according to claim 1.

5. The provisional model creation unit and the operational model creation unit are: A classification-type machine learning method (gradient boosting) is used to create the provisional model and the operational model. An anomaly prediction system according to claim 1.

6. The provisional model creation unit and the operational model creation unit are: The provisional model and the operational model are created using a classification-type machine learning method (logistic regression). An anomaly prediction system according to claim 1.

7. The provisional model creation unit and the operational model creation unit are: A classification-type deep learning method is used to create the provisional model and the operational model. An anomaly prediction system according to claim 1.

8. An abnormal occurrence prediction method for predicting abnormal phenomena occurring in an image forming apparatus, A market-operating machine data storage step involves collecting data from within the image forming apparatus at predetermined intervals and storing it for a certain period of time. A maintenance history data storage step for accumulating the maintenance history of each of the aforementioned image forming apparatuses, A training dataset creation step involves concatenating data that is considered equivalent to an abnormal situation based on the date of occurrence of the target abnormal phenomenon from the maintenance history with data randomly extracted from the maintenance history of image forming machines where the target abnormal phenomenon has not occurred, and generating a training dataset based on the set parameters. A preliminary model creation step involves creating a preliminary model that calculates the probability of the target anomaly occurring using the aforementioned training dataset, The following steps are performed to create a group: input a continuous dataset of the image forming apparatus over a certain period in the past, in the maintenance history data, into the provisional model; calculate the probability of the target abnormal phenomenon occurring each day; record the day on which the probability first exceeds a predetermined threshold for abnormal phenomenon occurrence and the nearest subsequent day of abnormal phenomenon occurrence; calculate the predicted number of days of occurrence leeway from the results; group the dataset according to the predicted number of days of occurrence leeway; change the parameters for each group; and create a new dataset for retraining. The operational model creation step involves formally creating an operational model for calculating the probability of the target anomaly occurring using the aforementioned retraining dataset, An anomaly prediction method including the occurrence of anomalies.

9. An abnormal occurrence prediction program that predicts abnormal phenomena occurring in an image forming apparatus in advance, A machine data storage function for machines in operation, which collects data from within the image forming apparatus at predetermined intervals and stores it for a certain period of time. A maintenance history data storage function for accumulating the maintenance history of each of the aforementioned image forming apparatuses, A training dataset creation function that combines data deemed to correspond to an abnormal situation based on the date of occurrence of the target abnormal phenomenon from the maintenance history with data randomly extracted from the maintenance history of image forming machines where the target abnormal phenomenon has not occurred, and generates a training dataset based on set parameters, A provisional model creation function that creates a provisional model for calculating the probability of the target anomaly occurring using the aforementioned training dataset, The provisional model is input with a continuous dataset of the image forming apparatus over a certain period in the past, based on the maintenance history data of the target abnormal phenomenon. The probability of the target abnormal phenomenon occurring each day is calculated, the day on which the probability first exceeds a predetermined threshold for abnormal phenomenon occurrence and the nearest subsequent day of occurrence are recorded, the predicted number of days of occurrence is calculated from the results, the dataset is grouped according to the predicted number of days of occurrence, the parameters are changed for each group, and a new dataset for retraining is created. An operational model creation function that formally creates an operational model for calculating the probability of the target anomaly occurring using the aforementioned retraining dataset, An anomaly prediction program that instructs a computer to perform the following actions.

Images