An artificial intelligence-based IT consumable inventory digital management method and system

Abstract

The application discloses an IT consumable inventory digital management method and system based on artificial intelligence, and relates to the technical field of inventory management. The method comprises the following steps: acquiring the work logs and replacement work orders of each IT consumable individual in an IT equipment cluster; extracting the cumulative printed page number, use intensity feature and idle feature for each IT consumable individual to form a feature snapshot updated over time; automatically constructing a partial order training sample using the inherent survival time sequence relationship between replacement events to label the high and low relationship of failure risk; training a risk comparison model composed of a deep neural network with weight sharing using the partial order training sample; inputting the feature snapshots of all in-stock and in-service consumables in the current inventory into the model to obtain a global risk score ranking, and accordingly preferentially recommending the consumable with the highest risk score as the out-of-stock object. The application realizes the ordering and differentiated scheduling of inventory consumables according to the failure urgency without relying on manual annotation of the remaining life of the consumables.

Description

Technical Field

[0001] This invention relates to the field of inventory management technology, and more specifically, to a method and system for digital management of IT consumables inventory based on artificial intelligence. Background Technology

[0002] In large enterprises and printing service centers, inventory management of IT consumables such as toner cartridges and drum units is a routine task in IT asset operations and maintenance. These consumables have a characteristic that differs from ordinary inventory: even individuals of the same specifications and from the same batch can have significantly different usable conditions. This is because each consumable experiences varying workloads during its service life. Some consistently handle high-coverage printing tasks, some frequently experience equipment start-ups and shutdowns, and some remain idle in the warehouse or equipment for extended periods. These differences result in varying toner conditions and drum aging levels, ultimately leading to a significant divergence in the remaining effective service life of different consumables.

[0003] Existing inventory management solutions do not address or utilize this issue. Chinese patent CN121961431A proposes an AI-based method for calculating inventory thresholds and purchase quantities, using a Long Short-Term Memory (LSTM) network to predict future demand trends and help determine when and how much to purchase. However, this method only addresses the total inventory level and does not differentiate between the specific conditions of individual consumables within the warehouse. Chinese patent CN120875750B proposes a warehouse management method for emergency supplies, adjusting storage conditions based on the nominal shelf life and environmental parameters. However, the actual failure of IT consumables is more driven by usage intensity than simply storage time; relying solely on the nominal shelf life is insufficient to accurately reflect the true condition of individual items.

[0004] Because existing technologies lack the means to perceive individual quality differences in consumables, two situations can easily occur in actual management: consumables that are close to being latently ineffective continue to remain in the warehouse because they are not discovered, and their quality deteriorates further until they are scrapped; while consumables that are still in good condition are replaced prematurely according to uniform rules (such as first in first out), and are not fully utilized.

[0005] Therefore, this paper proposes a digital management method and system for IT consumables inventory based on artificial intelligence to address the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an AI-based digital management method and system for IT consumables inventory, aiming to solve the problem that existing IT consumables inventory management methods cannot perceive the differences in the urgency of failure among individual consumables, resulting in a mismatch between outbound scheduling and the actual quality status of the consumables.

[0007] To achieve the above objectives, this invention provides a method for digital management of IT consumables inventory based on artificial intelligence, comprising the following steps:

[0008] S1. Obtain the job logs and replacement work orders for each IT consumable in the IT equipment cluster. The job logs include the number of pages printed, coverage, and timestamp for each print job. The replacement work orders include the consumable replacement time and the reason for replacement.

[0009] By continuously collecting the above two types of data, a digital record covering the entire process of individual consumables from warehousing and service to replacement is formed, providing a data foundation for subsequent feature extraction and model training.

[0010] S2. For each individual IT consumable, extract the cumulative number of printed pages, usage intensity characteristics, and idle characteristics based on its job log to form a feature snapshot that is updated over time.

[0011] Among the above features, the cumulative number of printed pages reflects the total load borne by the consumables, the usage intensity feature reflects the workload, and the idle feature reflects the potential for latent degradation due to prolonged inactivity. Transforming the original work records into a fixed-dimensional numerical vector provides a calculable and comparable basis for the states of different individual consumables.

[0012] Furthermore, the feature snapshot is constructed by encoding the cumulative number of printed pages, recent average coverage, daily average number of start-stops, and longest consecutive idle days, and applying a decay weighting to the encoded values ​​of recent average coverage and daily average number of start-stops according to a time window.

[0013] Since changes in usage patterns of consumables nearing their expiration date are often more indicative of risk than earlier changes, a decay-weighted approach is used to give recent data a larger proportion of the encoded values, while the influence of older data decreases over time. The weighted encoded values ​​are then concatenated in a fixed order to form a snapshot of the current features.

[0014] S3. Utilize the inherent survival time sequence relationship between replacement events to automatically construct partially ordered training samples from historical feature snapshots.

[0015] Each partially ordered training sample contains features of the first consumable and the second consumable, and is labeled that the failure risk of the first consumable is higher than that of the second consumable. The replacement event itself contains relative risk information about the consumables: under the same usage conditions, the risk state of a consumable that has been confirmed to have failed should be higher than the current state of a consumable that is still in normal service. By mining this temporal logic in the operation and maintenance data, supervisory signals can be obtained without relying on manual labeling of remaining lifespan.

[0016] Furthermore, the method for automatically constructing the partial order training samples is as follows: if the first consumable is replaced due to exhaustion or quality defects when the cumulative number of printed pages reaches the first page number, and the second consumable has accumulated more than the first page number under the same or more severe usage intensity as the first consumable and is still in normal service, then the feature snapshot recorded before the failure of the first consumable is used as the first feature, and the feature snapshot of the second consumable at the current moment is used as the second feature, to construct a partial order training sample that marks the failure risk of the first consumable as higher than that of the second consumable.

[0017] This method uses intensity criteria to ensure that both sides are in comparable service environments, avoiding the misjudgment of consumables that survive long-term low-intensity use as more durable. Daily replacement records are thus automatically converted into training sample pairs with clearly defined ordering relationships.

[0018] S4. Using the partially ordered training samples as supervision signals, train a risk comparison model composed of a deep neural network with shared weights, so that the risk score output by the model satisfies the partially ordered relation constraint.

[0019] The model's learning objective is to compare the relative risk of failure between any two consumables, rather than predicting their absolute remaining lifespan. If a consumable has a higher objective risk of failure, the model should assign it a higher score, thereby gaining the ability to distinguish the degree of urgency of failure between individual consumables.

[0020] Furthermore, the risk comparison model employs a Siamese network structure with shared weights in two input branches. The internal layers of the two branches are identical and their parameters are synchronized to ensure that both materials are evaluated according to the same scoring standard. The training loss function is an adaptive-interval ranking loss, determined based on the cumulative difference in the number of printed pages between the two consumables in the partially ordered training samples.

[0021] When the cumulative number of printed pages differs significantly, a larger interval is used to force the model to widen the score gap; when the difference is small, a smaller interval is used to allow the score gap to shrink accordingly, so that the model can distinguish consumables with similar risks.

[0022] S5. Input the current feature snapshots of all in-stock and in-service IT consumables in the current inventory into the trained risk comparison model to obtain a global risk score ranking.

[0023] "In stock" refers to consumables that have been purchased and stored but have not yet been issued or installed, while "in service" refers to consumables that have been installed on the printer and have not yet been replaced. All consumables in inventory are sorted in descending order of risk score; the higher the ranking, the greater the urgency of their failure.

[0024] S6. Perform inventory scheduling according to the global risk score, wherein the scheduling includes: when an outbound demand is generated, prioritizing the IT consumables with the highest risk score from the available inventory as the outbound object.

[0025] Unlike the traditional "first-in, first-out" rule, this strategy prioritizes consuming the inventory closest to its latent failure with each shipment, ensuring that high-risk consumables are used up before they become completely unusable, thus preventing them from deteriorating and becoming unusable due to long-term storage.

[0026] Furthermore, a dynamic threshold score is determined based on the risk score distribution of recently expired consumables. When calculating the effective inventory, only consumables with a risk score lower than the dynamic threshold score are included, and replenishment decisions are triggered based on the effective inventory and the average daily consumption.

[0027] The introduction of dynamic threshold scores means that inventory safety determination no longer depends on fixed number of days thresholds, but is dynamically linked to the actual quality status of consumables: only consumables with an acceptable health level are counted as effective inventory, and consumables with excessively high risk scores, although physically registered, are no longer considered as healthy inventory that can guarantee production.

[0028] Furthermore, after a new consumable replacement event occurs, the new event is transformed into a new partial-order training sample based on the survival time sequence relationship, and the risk comparison model is incrementally trained using the new partial-order training sample.

[0029] Through the aforementioned closed-loop mechanism, the model can continuously absorb the latest replacement data as the system operates, adapt to environmental changes such as equipment aging and the introduction of new consumable models, and ensure that the risk ranking capability remains consistent with the actual situation in the long term.

[0030] The present invention also provides an AI-based digital management system for IT consumables inventory, including a data acquisition module, a feature construction module, a sample generation module, a model training module, a risk ranking module, a scheduling execution module, and a feedback update module.

[0031] The data acquisition module acquires the operation logs and replacement work orders of each IT consumable in the IT equipment cluster. The operation logs include the number of pages printed, coverage, and timestamp for each print job, and the replacement work orders include the consumable replacement time and reason.

[0032] The feature construction module is connected to the data acquisition module, receives the job log, and extracts the cumulative number of printed pages, usage intensity features, and idle features for each individual IT consumable, forming a feature snapshot that updates over time, thus transforming discrete job records into a structured state representation.

[0033] The sample generation module is connected to the feature construction module, receives historical feature snapshots, and automatically constructs partial-order training samples by utilizing the inherent survival time sequence relationship between replacement events. Each partial-order training sample contains features of the first consumable and features of the second consumable, and marks that the failure risk of the first consumable is higher than that of the second consumable.

[0034] By mining the temporal logic hidden in the replacement events, this module continuously provides supervision signals for model training without the need for manual annotation.

[0035] Furthermore, the sample generation module specifically performs the following: if the first consumable is replaced due to exhaustion or quality defects when the cumulative number of printed pages reaches the first page number, and the second consumable has accumulated more than the first page number under the same or more severe usage intensity as the first consumable and is still in normal service, then the feature snapshot recorded before the failure of the first consumable is used as the first feature, and the feature snapshot of the second consumable at the current moment is used as the second feature, to construct a partially ordered training sample that indicates the failure risk of the first consumable is higher than that of the second consumable.

[0036] The model training module is connected to the sample generation module, receives the partially ordered training samples as supervision signals, and trains a risk comparison model composed of a deep neural network with shared weights, so that the risk score output by the model satisfies the partially ordered relation constraint.

[0037] Furthermore, the model training module uses a Siamese network structure with shared weights for two input branches to train the risk comparison model. The training loss function is an adaptive interval ranking loss, which is determined based on the difference in the cumulative number of printed pages between the two consumables in the partial order training samples. The larger the difference in the cumulative number of printed pages, the larger the interval value is used.

[0038] The risk ranking module is connected to the model training module and the feature construction module. It receives the trained risk comparison model and the current feature snapshots of all in-stock and in-service IT consumables in the current inventory, and outputs the global risk score ranking results to provide a quantitative basis for inventory scheduling.

[0039] The scheduling execution module is connected to the risk sorting module and receives the global risk score sorting result. When an outbound demand is generated, it prioritizes recommending the IT consumables with the highest risk score from the available inventory as the outbound object, so that the outbound order is consistent with the urgency of the consumables' failure.

[0040] The feedback update module is connected to the scheduling execution module and the model training module respectively. After a new consumable replacement event occurs, the new event is transformed into a new partial order training sample according to the survival time sequence relationship, and the model training module is triggered to perform incremental training using the new partial order training sample, so that the model can be continuously updated as the system runs.

[0041] The technical effects and advantages of this invention are as follows:

[0042] (1) This method extracts the cumulative number of printed pages, usage intensity features, and idle features of each consumable from the printer job log and consumable replacement work order. After time decay weighting and other encoding processes, a feature snapshot that is updated over time is formed, thereby transforming the actual service experience of the consumable into a numerical vector that can be calculated and compared. The supervision signal required for training the model comes directly from the order of replacement events: the feature snapshot of a consumable that has been confirmed to have failed before its failure is paired with the current feature snapshot of a still-in-service consumable with similar usage intensity but a larger cumulative print volume, and the former is marked as having a higher risk. This process does not require manual labeling of the remaining lifespan of any consumable.

[0043] (2) The risk comparison model is trained using a shared-weights Siamese network structure. The training objective is an adaptive-interval ranking loss, with the interval determined by the difference in the cumulative number of printed pages between the two consumables in the training pair. After training, the model is input with a feature snapshot of any consumable, and it outputs a risk score. By arranging all consumables in the inventory from highest to lowest score, the order of failure urgency is obtained. In this way, inventory management no longer focuses on quantity or inbound time, but allows us to see the relative status of each individual consumable.

[0044] (3) Outbound scheduling follows a "failure priority" approach. Whenever a replacement is needed, the consumable with the highest risk score in the warehouse is prioritized, ensuring that the inventory closest to latent failure is used first. Replenishment decisions no longer rely on fixed-day thresholds but instead determine a dynamic threshold based on the recent distribution of risk scores for actual failed consumables. Only consumables with scores below this threshold are counted as valid inventory. The entire solution draws its data from existing printer operation logs and replacement work orders, eliminating the need for additional sensors and facilitating direct implementation within the existing IT operations environment. Attached Figure Description

[0045] Figure 1 This is a system module framework diagram of the present invention.

[0046] Figure 2 This is a flowchart illustrating the implementation of the method of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] As attached Figures 1 to 2 The present invention discloses an AI-based digital management method and system for IT consumables inventory, used to manage an IT equipment cluster consisting of a large number of printers.

[0050] In this scenario, individual IT consumables refer to the toner cartridges or drum units used by each printer, and each toner cartridge or drum unit is assigned a unique identifier. This identifier is linked to the consumable through a barcode or RFID tag and is used throughout the entire process of the consumable's warehousing, service, replacement, and eventual disposal.

[0051] S1. Obtain the operation logs and replacement work orders of each IT consumable in the IT equipment cluster.

[0052] First, obtain the job logs and replacement work orders for each IT consumable item in the IT equipment cluster. The job logs include the number of pages printed, coverage, and timestamp for each print job. The replacement work orders include the consumable replacement time and reason.

[0053] Regarding the method of obtaining job logs, in one implementation, a data acquisition program is deployed on the management server to periodically poll each printer using a standard simple network management protocol. For example, the polling interval is set to 5 minutes. Each poll reads the printer's total page count and coverage statistics. The difference in the total page count between two consecutive polls is the number of pages printed in that period. This number of printed pages is then associated with the coverage rate, the timestamp of the polling time, and the consumable identifier currently installed on the printer to form a job log record.

[0054] In another implementation, a data reporting agent is embedded in the printer firmware. The agent immediately encapsulates the number of pages printed, coverage percentage, and completion timestamp into a data packet and proactively sends it to the data acquisition server. The server then parses the packet to generate a job log. Compared to polling, this method can more accurately record the actual time of each print job.

[0055] The method for obtaining a replacement work order is as follows:

[0056] The system establishes an interface connection with the IT operations and maintenance management system's database, periodically retrieving closed replacement work orders from it. Each replacement work order includes a unique identifier for the replaced consumable, a timestamp indicating the replacement operation was completed, and a replacement reason code.

[0057] The replacement reason code is standardized into three types:

[0058] "Exhausted" means that the consumables have been replaced normally because the print quality has deteriorated to an unacceptable level.

[0059] "Quality defects" means that consumables are replaced prematurely due to abnormal reasons such as scratches on the photosensitive drum or toner clumping.

[0060] "Manual removal" means that the consumables themselves still have a usable lifespan but are removed due to management factors such as equipment allocation.

[0061] In subsequent model training, only the replacement events corresponding to the first two reasons are considered as real failure events, because both of these situations clearly indicate that the consumables have lost or severely reduced their normal service capability.

[0062] S2. For each individual IT consumable, construct a feature snapshot based on its job log.

[0063] Next, for each individual IT consumable, a feature snapshot is constructed based on its operation log. The feature snapshot is a fixed-dimensional numerical vector used to characterize the usage status of that consumable at a specific moment.

[0064] For any individual IT consumable, the system aggregates and statistically analyzes all job logs within the time window from the moment it is installed in the printer and begins service to the current system time, extracting the following features.

[0065] First, calculate the cumulative number of pages printed. This value reflects the total physical load that the consumable has undertaken, and it is the sum of the number of pages printed in all job logs during its service life.

[0066] Secondly, usage intensity features are extracted. In one implementation, usage intensity features include recent average coverage, which is the arithmetic mean of the coverage recorded by the most recent m print jobs before the current time, where m is a preset positive integer, for example, 50.

[0067] Higher coverage means greater consumption of toner and drum per page. Therefore, recent average coverage can reflect the average workload that consumables have recently endured.

[0068] In another implementation, the intensity characteristic also includes the average number of daily start-stop cycles, which is the total number of print jobs during the service period divided by the total number of days in service. Each start-up and shutdown of the printer causes additional mechanical impact to components such as the photosensitive drum, so the frequency of start-stop cycles is another important dimension for measuring the wear intensity of consumables.

[0069] Secondly, extract idle characteristics. Even when not in use, IT consumables can experience slow, subtle quality degradation, such as toner clumping due to ambient moisture. This subtle degradation is highly correlated with the continuous period of inactivity the consumables have undergone.

[0070] The system extracts the longest consecutive idle days as the idle feature. The calculation method is as follows: sort the timestamps of all operation logs during the service period of the consumable in ascending order, calculate the number of days between every two adjacent timestamps, and take the maximum value.

[0071] After extracting the original feature values, the system encodes them to integrate these feature values ​​of different dimensions and scales into a compact numerical vector.

[0072] A time-decay weighted strategy is preferred in the processing. Specifically, the time period before the observation time is divided into w equal-length time windows, and the average coverage value in the k-th window is multiplied by... As a weighted contribution to this window. It is a preset constant that satisfies k starts counting from 0, and k=0 corresponds to the most recent time window.

[0073] The effect of this processing is that the most recent usage data carries the largest weight in the encoded value, while older usage history has a smaller impact on the current encoded value. This strategy is adopted because changes in usage patterns of consumables nearing expiration are often more indicative of risk than earlier usage patterns. If the average daily start-stop count is also extracted, its historical sequence is also subjected to decay weighting in the same way.

[0074] As another implementation method, linear decay weighting can also be used, that is, the weight of the k-th window is... .

[0075] As another implementation method, when historical data is considered to have reference value, the attenuation weighting step can be omitted, and the arithmetic mean of the coverage of the most recent m tasks can be directly taken as the encoding value.

[0076] Finally, the encoded values ​​of the cumulative number of printed pages, the attenuated weighted coverage rate, the average number of daily start-stop cycles, and the longest consecutive idle days are concatenated in a fixed order to form a vector with fixed dimensions. The vector This is a feature snapshot.

[0077] To unify the numerical scale of each dimension, min-max normalization was performed on each dimension, mapping the values ​​to the [0, 1] interval. The original numerical range of the cumulative number of printed pages differed significantly from that of the coverage rate. After normalization, each dimension fell within the [0, 1] interval, avoiding the impact of scale imbalance on subsequent model training.

[0078] Feature snapshots are not static. As new print jobs are generated, each time the system needs to perform a risk assessment on the consumables, a new feature snapshot is generated based on the latest job log at that time. Each newly generated feature snapshot is appended with a generation timestamp and written to the feature snapshot storage area, accumulating a chronologically ordered state sequence for that consumable. This state sequence will serve as the source of raw materials for constructing training samples in step S3.

[0079] S3. Automatically construct partially ordered training samples by utilizing the inherent survival time relationships between replacement events.

[0080] Then, by leveraging the inherent time-series relationships between replacement events, partially ordered training samples are automatically constructed from historical feature snapshots. Each partially ordered training sample contains features of the first consumable and the second consumable, and is labeled that the failure risk of the first consumable is higher than that of the second consumable.

[0081] The construction process consists of the following steps.

[0082] First, identify the failed samples and extract their pre-failure states. The sample generation module iterates through all replacement work orders in the database, filters out work orders with the replacement reason as "exhausted" or "quality defect", and marks the consumables associated with these work orders as failed samples.

[0083] One of the failed samples is designated as consumable A. The system obtains the cumulative number of printed pages when consumable A is replaced, denoted as . Simultaneously, retrieve all historical feature snapshots of consumable A from the feature snapshot storage area, sort them by timestamp, and extract the snapshot closest to the replacement time with a timestamp earlier than the replacement time, denoted as... . This is the last snapshot of the state of consumable A before it fails, and it is a numerical representation of the failure state.

[0084] Secondly, the system searches for comparable surviving samples and verifies the preconditions. It searches the consumables database for consumables of the same printer model as consumable A that are currently in use. For a candidate surviving consumable B, the system reads its current cumulative number of printed pages. and current feature snapshot .

[0085] Before pairing consumable A and consumable B as training sample pairs, the system must first perform a usage intensity condition determination. The purpose of this determination is to ensure the fairness of the comparison. For example, if consumable A fails prematurely due to prolonged operation under harsh conditions of high coverage and frequent start-stop cycles, while consumable B has survived to this day by operating easily under low intensity conditions, then B's longer survival does not necessarily mean that B is more durable, and their risk states are not comparable.

[0086] In one implementation, the intensity condition determination is accomplished by simultaneously comparing the daily average number of start-stop cycles s and the recent average coverage rate c of the two consumables.

[0087] Only when and Only when the usage intensity of consumable B is deemed to be no less than that of consumable A is it considered that the usage intensity of consumable B is not less than that of consumable A. For example, consumable A has an average daily start-stop frequency of 8 times and a recent average coverage rate of 5%, while consumable B has corresponding values ​​of 10 times and 6%, respectively. Since both dimensions are not lower than those of consumable A, the judgment condition is met. In another implementation, preset weights can be assigned to s and c respectively, and a comprehensive intensity score can be calculated. When the comprehensive score of consumable B is greater than or equal to the comprehensive score of consumable A, the judgment condition is met.

[0088] When the strength condition is met, and When consumable B, under comparable or more stringent conditions, bears a higher cumulative printing volume, the system constructs a partially ordered training sample. This sample is based on... As the first feature, with The second feature indicates that the consumables corresponding to the first feature have a higher failure risk than the consumables corresponding to the second feature.

[0089] Through this mechanism, replacement records generated during routine operations and maintenance are automatically transformed into training sample pairs with clearly defined risk ranking relationships. The entire process fully utilizes the inherent temporal logic of the operations and maintenance data itself, requiring no additional annotation effort.

[0090] S4. Using partially ordered training samples as supervision signals, train a risk comparison model composed of a deep neural network with shared weights.

[0091] Then, using the partially ordered training samples as supervision signals, a risk comparison model composed of a deep neural network with shared weights is trained so that the risk score output by the model satisfies the partially ordered relation constraint.

[0092] The model receives a feature snapshot. As input, the output is a scalar value, denoted as the risk score. . The higher the score, the more urgent the risk of failure. For any two consumables, if one of them has a higher risk of failure, then the model should give the former a higher score.

[0093] To achieve this goal, the model employs a Siamese network structure with shared weights in both input branches. The internal layers of the two branches are identical, and they share all network parameters.

[0094] In one implementation of the branch internal structure, each branch consists of three fully connected layers stacked sequentially. For example, if the input feature dimension is 64, the first fully connected layer outputs 32 dimensions, the second fully connected layer outputs 16 dimensions, and the third fully connected layer outputs a single 1-dimensional value, which is the risk score output for that branch. Each layer is followed by a batch normalization layer and a modified linear unit activation function.

[0095] During training, the first feature of a partially ordered training sample is... Feed into a branch, second feature Send it to another branch. and Each branch propagates forward layer by layer within its own branch for calculation, and each branch outputs a risk score. and .

[0096] The training loss function is an adaptive-margin ranking loss. Its core idea is: for each partially ordered training pair, the score of the high-risk sample is required to be... The score must be greater than that of low-risk samples. Furthermore, the difference between the two must be at least a certain distance. Loss value The calculation method is as follows .

[0097] It should be noted that the interval It is not a fixed constant that applies to all samples, but rather is adaptively determined for each pair of samples. Its setting is based on the cumulative difference in the number of pages printed by the two consumables. .

[0098] when A large value indicates that one has already failed while the other is far from the end of its lifespan; the risks between them should be significantly different, hence the setting of a large value. This forces the model to create a sufficiently large score gap on these types of samples.

[0099] when When the value is smaller, it indicates that the two consumables are already in very similar condition, with subtle differences in risk; therefore, a smaller value is used. This allows for a smaller difference in the scores output by the model.

[0100] In one implementation, and It exhibits a linear relationship, that is , and The default value is a positive number. For example, Take 0.1, Take 0.01, when Page time ,when Page time The greater the difference, the larger the interval. As another implementation method, ,in This refers to the nominal maximum number of pages that can be printed for the same type of consumable. This is a preset scaling factor.

[0101] Training uses the Adam optimizer to automatically update network parameters. Training stops when the loss value no longer decreases significantly after several consecutive training epochs, and the model parameters at this point are saved for subsequent steps.

[0102] S5. Input the current feature snapshots of all in-stock and active IT consumables in the current inventory into the model to obtain a global risk score ranking.

[0103] After the model is trained, the current feature snapshots of all in-stock and in-service IT consumables in the current inventory are input into the trained risk comparison model to obtain a global risk score ranking.

[0104] Each time this step is performed, the scope of the assessment is first determined. The system queries the inventory database for all IT consumables marked as "In Stock" or "In Service". "In Stock" means the consumable has been purchased and stored but has not yet been issued and installed, while "In Service" means the consumable has been installed on the printer and has not yet been replaced.

[0105] For each individual consumable within the scope, the system generates a current feature snapshot based on its latest job log up to the current moment, following the feature construction process described in S2. This feature snapshot is then fed into a risk comparison model loaded with the model parameters saved in S4. After one forward calculation, the model outputs the corresponding risk score.

[0106] Once the risk scores for all consumables to be evaluated have been calculated, the system binds the unique identifier of each consumable to its risk score value, and sorts them in descending order of risk score, generating a global risk score ranking list. The top of the list represents the consumables with the most urgent failure risk in the entire IT equipment cluster. Simultaneously, the system writes the risk score value of each consumable back to the corresponding record field in the inventory database for later querying in subsequent scheduling steps.

[0107] S6. Perform inventory scheduling based on global risk scores.

[0108] Finally, inventory scheduling is performed based on the global risk score ranking. Scheduling includes: when an outbound demand arises, prioritizing the IT consumables with the highest risk score from available inventory as the outbound target.

[0109] In terms of outbound scheduling, an outbound request is triggered when a printer reports that its consumables are out of stock, or when maintenance personnel initiate a replacement request through the work order system. After receiving the outbound request, the scheduling execution module first identifies the type of consumables required by the request.

[0110] Then, the system filters all consumables of that model that are "in stock" from the global risk score ranking list. The system pushes the consumable's unique identifier along with its storage location information to the warehouse management terminal or the mobile terminal of maintenance personnel, prioritizing its release. This "expiration-first release" strategy is continuously implemented, ensuring that consumables whose quality deteriorates slowly due to prolonged storage are used before they become completely unusable.

[0111] In terms of replenishment decisions, this method uses a dynamic critical point mechanism to replace the traditional fixed number of days threshold.

[0112] First, the dynamic critical score is determined. The scheduling execution module periodically reviews failure records from a past period, such as the past 30 days. It collects the risk score values ​​of all consumables replaced due to "exhaustion" or "quality defects" during this period, calculated last time before failure, forming a reference sample set.

[0113] In one implementation, the system takes the lower quartile of all risk scores in the reference sample set as the current dynamic threshold score. For example, if the reference sample set has 100 risk scores, the 25th value after sorting in ascending order is taken as the dynamic threshold score.

[0114] This critical score provides an adaptive benchmark: if the current risk score of an in-stock consumable exceeds this threshold, it indicates that the consumable's health status has entered the risk score distribution range of the actual failed samples and can be considered quasi-failed inventory. Alternatively, the system can take the arithmetic mean of the risk scores in the reference sample set minus one standard deviation as the dynamic critical score.

[0115] Next, the system calculates the effective inventory. It iterates through all consumables in the "in stock" state and reads their current risk score recorded in the inventory database. Only consumables with a risk score lower than the dynamic threshold are included in the effective inventory. Those consumables in stock with a risk score higher than the threshold, although physically still stored on warehouse shelves, are no longer considered healthy inventory to ensure production from a management perspective. For example, a toner cartridge that has been stored in the warehouse for six months and has a risk score exceeding the threshold, although physically recorded, is excluded from the effective inventory by the system.

[0116] Finally, the system determines whether to trigger replenishment. It calculates the average daily consumption over a past period, which is the total consumption of the same type of consumables during that period divided by the number of days. Then, it divides the current effective inventory by the average daily consumption to obtain the estimated number of days the current effective inventory can sustain the stock. When the estimated number of days is lower than a preset safety threshold, the system automatically generates a replenishment trigger signal, sending it along with key information such as the current effective inventory and average daily consumption to the procurement management module, reminding relevant personnel to promptly replenish the stock.

[0117] This concludes the description of a complete closed-loop operation cycle, from data collection and status feature perception to risk comparison and ranking and differentiated scheduling execution.

[0118] Based on the aforementioned closed-loop operation, this method also possesses continuous self-evolution capabilities. Whenever the system records a new replacement event with the reason of "exhaustion" or "quality defect," the feedback update module is activated.

[0119] Using this newly failed consumable as a reference, this module selects individuals that meet the conditions from consumables of the same model that are currently in service, according to the same partial order relation construction rules as in S3, and automatically generates several new partial order training samples, which are then stored in the training sample buffer.

[0120] When the number of newly added samples in the buffer reaches a preset number, or when a preset time period has elapsed since the last incremental training, the system triggers an incremental training. Incremental training starts with the parameters of the current model trained in S4, and only uses this batch of newly added samples to update the parameters in a few rounds. The learning rate is set to a small proportion of the initial training learning rate, such as 0.1 times.

[0121] After incremental training is completed, the model parameters are replaced with the updated version. Through this mechanism, the risk comparison model can continuously adapt to subtle environmental changes brought about by equipment aging, the introduction of new types of consumables, etc., so that the risk ranking ability can remain accurate over a long period of time.

[0122] The implementation process of the method of the present invention has been described in detail above. Based on the same inventive concept, the present invention also provides an artificial intelligence-based digital inventory management system for IT consumables.

[0123] The system includes a data acquisition module, which is responsible for acquiring the operation logs and replacement work orders of each IT consumable through polling or proxy reporting. Its specific operation is as described in S1 above.

[0124] The feature construction module is connected to the data acquisition module, receives job logs, constructs a feature snapshot for each individual IT consumable and writes it to the storage area, and its specific operation is as described in S2 above.

[0125] The sample generation module is connected to the feature construction module. It reads historical snapshots from the feature snapshot storage area and automatically constructs partial-order training samples by utilizing the survival time sequence relationship between replacement events. The specific operation is as described in S3 above.

[0126] The model training module is connected to the sample generation module, receives partially ordered training samples as supervision signals, and trains a shared weighted Siamese deep neural network as a risk comparison model. Its specific operation is as described in S4 above.

[0127] The risk ranking module is connected to the model training module and the feature construction module. It loads the trained model parameters, inputs the feature snapshots of all currently stocked and in-service consumables into the model, outputs a global risk score ranking list, and writes the scores back to the inventory database. The specific operation is as described in S5 above.

[0128] The scheduling execution module is connected to the risk ranking module. Based on the ranking results, it prioritizes the release of high-risk consumables and uses a dynamic critical point mechanism to count effective inventory and trigger replenishment. The specific operation is as described in S6 above.

[0129] The feedback update module is connected to the scheduling execution module and the model training module respectively. After a new failure replacement event occurs, it generates new partial order training samples and triggers the model training module to perform incremental training. Its specific operation is as described in S6 above.

[0130] Regarding the system deployment method, in one implementation, all modules are deployed on the same server, and data is transferred between modules through in-memory data structures and function calls.

[0131] In another implementation, the data acquisition module is deployed on an edge computing gateway close to the printer, collecting printer data locally via the local area network to reduce latency; the remaining modules are deployed on a cloud server, with data transmission between the edge gateway and the cloud via a message queue. Regardless of the deployment method, the feedback update module runs in the cloud, triggering the generation of incremental samples and online updates of the model by listening for change event notifications from the scheduling execution module or periodically polling the work order database.

[0132] The foregoing has provided a detailed description of the AI-based digital inventory management method and system for IT consumables provided by this invention. For those skilled in the art, various modifications and combinations can be made to the above embodiments without departing from the concept of this invention, and these modifications and combinations should also be considered within the scope of protection of this invention.

Claims

1. A method for digital management of IT consumables inventory based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain the operation logs and replacement work orders of each IT consumable in the IT equipment cluster. The operation logs include the number of pages printed, coverage, and timestamp for each print job. The replacement work orders include the consumable replacement time and the reason for replacement. S2. For each individual IT consumable, extract the cumulative number of printed pages, usage intensity characteristics, and idle characteristics based on its job log to form a feature snapshot that is updated over time. S3. Utilize the inherent survival time sequence relationship between replacement events to automatically construct partial order training samples from historical feature snapshots. Each partial order training sample contains the features of the first consumable and the features of the second consumable, and marks that the failure risk of the first consumable is higher than that of the second consumable. S4. Using the partially ordered training samples as supervision signals, train a risk comparison model composed of a deep neural network with shared weights so that the risk score output by the model satisfies the partially ordered relation constraint. S5. Input the current feature snapshots of all in-stock and in-service IT consumables in the current inventory into the trained risk comparison model to obtain a global risk score ranking. S6. Perform inventory scheduling according to the global risk score, wherein the scheduling includes: when an outbound demand is generated, prioritizing the IT consumables with the highest risk score from the available inventory as the outbound object.

2. The method for digital management of IT consumables inventory based on artificial intelligence according to claim 1, characterized in that, The method for automatically constructing partially ordered training samples in step S3 is as follows: If the first consumable is replaced due to exhaustion or quality defects when the cumulative number of printed pages reaches the first page count, and the second consumable has accumulated more printed pages than the first consumable under the same or more stringent usage intensity and is still in normal working condition, then... Then, the feature snapshot of the last record before the failure of the first consumable is used as the first feature, and the feature snapshot of the current moment of the second consumable is used as the second feature to construct a partially ordered training sample that marks the failure risk of the first consumable as higher than that of the second consumable.

3. The method for digital management of IT consumables inventory based on artificial intelligence according to claim 2, characterized in that, The feature snapshot described in step S2 is constructed as follows: The cumulative number of printed pages, recent average coverage, daily average number of start-stops, and longest consecutive idle days are encoded separately. The encoded values ​​of recent average coverage and daily average number of start-stops are weighted by attenuation according to the time window, and the weighted encoded values ​​are spliced ​​in a fixed order to form the current feature snapshot.

4. The method for digital management of IT consumables inventory based on artificial intelligence according to claim 3, characterized in that, The risk comparison model described in step S4 adopts a Siamese network structure with two input branches sharing weights. The training loss function is an adaptive interval ranking loss, which is determined based on the difference in the cumulative number of printed pages between the two consumables in the partial order training samples.

5. The method for digital management of IT consumables inventory based on artificial intelligence according to claim 1, characterized in that, Step S6, which involves sorting inventory based on global risk scores, further includes: A dynamic threshold score is determined based on the risk score distribution of recently expired consumables. When calculating the effective inventory, only consumables with a risk score lower than the dynamic threshold score are included, and replenishment decisions are triggered based on the effective inventory and the average daily consumption.

6. The method for digital management of IT consumables inventory based on artificial intelligence according to claim 5, characterized in that, After a new consumable replacement event occurs, the new event is transformed into a new partial-order training sample based on the survival time sequence relationship, and the risk comparison model is incrementally trained using the new partial-order training sample.

7. An AI-based digital inventory management system for IT consumables, characterized in that, include: The data acquisition module acquires the operation logs and replacement work orders of each IT consumable in the IT equipment cluster. The operation logs include the number of pages printed, coverage, and timestamp for each print job, and the replacement work orders include the consumable replacement time and reason. The feature construction module is connected to the data acquisition module, receives the job log, and extracts the cumulative number of printed pages, usage intensity features, and idle features for each individual IT consumable, forming a feature snapshot that updates over time. The sample generation module is connected to the feature construction module, receives historical feature snapshots, and automatically constructs partial-order training samples by utilizing the inherent survival time sequence relationship between replacement events. Each partial-order training sample contains the features of the first consumable and the features of the second consumable, and marks that the failure risk of the first consumable is higher than that of the second consumable. The model training module is connected to the sample generation module, receives the partial order training samples as supervision signals, and trains a risk comparison model composed of a deep neural network with shared weights, so that the risk score output by the model satisfies the partial order relation constraint. The risk ranking module is connected to the model training module and the feature construction module. It receives the trained risk comparison model and the current feature snapshots of all in-stock and in-service IT consumables in the current inventory, and outputs the global risk score ranking result. The scheduling and execution module is connected to the risk ranking module, receives the global risk score ranking result, and when an outbound demand is generated, prioritizes recommending the IT consumables with the highest risk score from the available inventory as the outbound object.

8. The AI-based digital inventory management system for IT consumables according to claim 7, characterized in that, The sample generation module specifically performs the following: If the first consumable is replaced due to exhaustion or quality defects when the cumulative number of printed pages reaches the first page count, and the second consumable has accumulated more printed pages than the first consumable under the same or more stringent usage intensity and is still in normal working condition, then... Then, the feature snapshot of the last record before the failure of the first consumable is used as the first feature, and the feature snapshot of the current moment of the second consumable is used as the second feature to construct a partially ordered training sample that marks the failure risk of the first consumable as higher than that of the second consumable.

9. The AI-based digital inventory management system for IT consumables according to claim 8, characterized in that, The model training module uses a Siamese network structure with two input branches sharing weights to train the risk comparison model. The training loss function is an adaptive interval ranking loss, which is determined based on the difference in the cumulative number of printed pages between the two consumables in the partial order training samples.

10. The AI-based digital inventory management system for IT consumables according to claim 7, characterized in that, It also includes a feedback update module, which is connected to the scheduling execution module and the model training module respectively. After a new consumable replacement event occurs, the new event is transformed into a new partial order training sample according to the survival time sequence relationship, and the model training module is triggered to perform incremental training using the new partial order training sample.

Images