Local and global redundancy optimization repair method and system based on data mining

Abstract

The invention relates to the technical field of memory fault repair, and particularly discloses a local and global redundancy optimization repair method and system based on data mining, and the method comprises the steps: collecting historical fault data of a memory, cleaning the historical fault data, and extracting fault bit features; carrying out clustering analysis on the extracted fault bit features by adopting a clustering algorithm, and dividing a high-frequency fault region of the fault bit; constructing a distance model of the fault bit and the redundant unit, wherein the distance model fuses the physical distance and the logic distance of the fault bit and the redundant unit; when a new fault bit is detected, preferentially allocating a local redundancy unit with a closer logic distance and / or physical distance based on the distance model for repairing, and updating a resource occupation state of a local redundancy pool; and if the local redundant unit of the area where the new fault bit is located is exhausted, selecting an optimal resource from the global redundant unit according to the distance weight based on the distance model to repair.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of memory fault repair, in particular to a local and global redundancy optimization repair method and system based on data mining. BACKGROUND

[0002] With the evolution of integrated circuit technology to deep sub-micron, nanometer scale, high-density memory (such as static random access memory, DRAM, NAND Flash) has become the core memory unit of modern electronic systems. Due to the extremely high device density and stringent process requirements of the memory, small process fluctuations, material defects in the manufacturing process or aging effects during use can easily cause storage unit (bit / cell) failure, directly affecting chip yield, system stability and reliability.

[0003] To solve the problem of memory failure, the prior art usually adopts a redundancy repair scheme, that is, by setting local redundancy (Local Redundancy) and global redundancy (Global Redundancy) resources, to replace the failed storage unit. However, the prior art has the problem that when a fixed redundancy allocation rule is used, it cannot adapt to the dynamically changing fault distribution, resulting in insufficient redundancy resources in some areas and idle resources in some areas; and the location correlation between the fault unit and the redundancy unit is not considered, and the blind use of global redundancy can easily introduce signal delay and interference, reducing the repair effect. SUMMARY

[0004] The purpose of the present application is to provide a local and global redundancy optimization repair method and system based on data mining to solve the above technical problems.

[0005] The purpose of the present application can be achieved by the following technical solutions: A local and global redundancy optimization repair method and system based on data mining, comprising the following steps: Step S1: Collecting the historical fault data of the memory, after cleaning the historical fault data, extracting the fault bit features; the historical fault data includes fault address, fault type and environmental parameters; the fault bit features include fault bit coordinates, belonging storage block and timing characteristics; Step S2: Clustering analysis of the extracted fault bit features using a clustering algorithm, dividing the high-frequency fault area of the fault bit; Step S3: Constructing a distance model of the fault bit and the redundancy unit, the distance model fuses the physical distance and the logical distance of the fault bit and the redundancy unit; the physical distance is the physical location proximity of the redundancy unit and the fault unit within the chip, and the logical distance is an abstract distance based on the fault pattern, established by association rule mining; Step S4: when a new fault bit is detected, a local redundancy unit with a shorter logical distance and / or physical distance is preferentially assigned for repair based on the distance model, and the resource occupation state of the local redundancy pool is updated; Step S5: if the local redundancy unit in the region where the new fault bit is located is exhausted, the optimal resource is selected from the global redundancy unit for repair based on the distance model according to the distance weight.

[0006] As a further scheme of the present application, the step S2 specifically comprises: S2.1: from the fault bit features extracted in step S1, core features related to strong fault region aggregation are screened, the screening method is to retain the fault bit coordinates and the storage block to which the fault bit belongs, the timing characteristics are converted into auxiliary screening conditions, and if the feature dimension is higher than a preset threshold, a dimension reduction algorithm is used to compress the dimension of the features; S2.2: a clustering algorithm is selected according to the distribution characteristics of the memory fault bits, and the selected clustering algorithm is parameterized; S2.3: the feature data optimized in step S2.1 is input into the initialized clustering algorithm, and clustering operation is performed to generate a plurality of fault bit clusters, invalid clusters with a size smaller than a preset threshold are removed, and valid clusters are retained; S2.4: the core attribute of each valid cluster is calculated, the high-frequency fault region is determined according to the core attribute, the region boundary is determined and stored, and a high-frequency fault region index table is formed.

[0007] As a further scheme of the present application, the step S3 specifically comprises the following contents: S3.1: a physical distance calculation rule is defined, the physical position parameters of the redundancy unit and all storage units are extracted based on the physical design layout of the memory chip, the physical position proximity of the redundancy unit and the fault unit is calculated by using the Euclidean distance or Manhattan distance formula, the physical distance quantization value is obtained after standardization processing, and at the same time, the redundancy unit across the physical partition is set with a physical distance penalty coefficient to correct the calculation result in combination with the chip wiring constraint; S3.2: a logical distance model is constructed by association rule mining; S3.3: a comprehensive distance model is constructed by fusing the physical distance and the logical distance; S3.4: a redundancy unit distance index table is established, the comprehensive distance calculation results of all redundancy units after classification of the local redundancy pool and the global redundancy pool are stored, and a distance index table indexed by the fault bit region and the fault mode is constructed.

[0008] As a further scheme of the present application, the step S3.2 specifically comprises the following contents: S3.2.1: Based on the historical fault data after cleaning in step S1, the association between the fault bit features and the fault modes is extracted to form an association analysis data set; S3.2.2: The association rules mining is performed on the association analysis data set by using the Apriori algorithm or the FP-Growth algorithm, and strong association rules with confidence and support higher than preset threshold values are screened out; S3.2.3: Based on the strong association rules, a logical distance quantification standard is defined, the matching condition of the fault mode of the fault bit and the redundant unit adaptation mode is given a corresponding logical distance value, and the logical distance is standardized to a preset interval.

[0009] As a further scheme of the application, the step S3.3 specifically includes the following contents: S3.3.1: The weight coefficients of the physical distance and the logical distance are determined by using the analytic hierarchy process or machine learning modeling based on the historical repair success rate, and the weight distribution takes into account the repair efficiency and the repair stability; S3.3.2: The comprehensive distance between the fault bit and the redundant unit is calculated by using a weighted summation formula, and the weighted summation formula is D=ω1×physical distance+ω2×logical distance, wherein ω1 is the weight coefficient of the physical distance, ω2 is the weight coefficient of the logical distance, and ω1+ω2=1; S3.3.3: The matching degree of the comprehensive distance model is verified based on the historical fault repair data, if the matching degree is lower than a preset threshold value, the weight coefficient or the threshold parameter of the association rules mining is adjusted, and the iteration optimization is performed until the model meets the preset accuracy requirement.

[0010] As a further scheme of the application, the step S4 specifically includes: S4.1: The fault address, the fault type, the belonging storage block, the real-time environment parameter and the fault triggering time sequence of the new fault bit are collected, the core features of the new fault bit are extracted from the collected data based on preset fault bit feature extraction rules to form a feature vector, and the core features include the fault bit coordinate, the belonging storage block and the time sequence characteristics; S4.2: The belonging storage block and the fault bit coordinate of the new fault bit are matched with the high-frequency fault area index table to determine whether the new fault bit is located in the high-frequency fault area, and the corresponding exclusive local redundancy pool is positioned according to the physical position of the new fault bit, and the local redundancy pool is pre-divided according to the memory physical partition; S4.3: The preset redundancy unit distance index table is called to take the belonging area and the fault mode of the new fault bit as indexes to query the comprehensive distances between all available redundancy units in the local redundancy pool and the new fault bit, the comprehensive distance is the weighted summation result of the physical distance and the logical distance, the first N candidate local redundancy units are screened out in order of the comprehensive distance from small to large, and N is a preset candidate number; S4.4: Perform working state detection, adaptation mode matching verification and physical compatibility verification on the candidate local redundancy unit, select the local redundancy unit with the best adaptation, and start the fault repair process, the repair process including redundancy unit activation, fault bit data migration and address mapping update; S4.5: After the repair is completed, mark the used local redundancy unit as "occupied" state and remove it from the available resource list, synchronously update the associated record in the redundancy unit distance index table, and if the remaining available resources of the local redundancy pool are lower than the preset warning threshold, output a resource warning signal.

[0011] A local and global redundancy optimization repair system based on data mining, characterized by comprising: A data acquisition module: collects historical fault data of the memory, extracts fault bit features after cleaning the historical fault data, the historical fault data including fault address, fault type and environmental parameters, and the fault bit features including fault bit coordinates, belonging memory block and timing characteristics; A clustering analysis module: performs clustering analysis on the extracted fault bit features using a clustering algorithm, and divides the high-frequency fault area of the fault bit; A model construction module: constructs a distance model of the fault bit and the redundancy unit, the distance model fusing the physical distance and the logical distance of the fault bit and the redundancy unit, the physical distance being the physical position proximity of the redundancy unit and the fault unit in the chip, and the logical distance being an abstract distance based on the fault mode and established through association rule mining; A first repair module: when a new fault bit is detected, preferentially allocates a local redundancy unit with closer logical distance and / or physical distance for repair based on the distance model, and updates the resource occupation state of the local redundancy pool; A second repair module: if the local redundancy unit in the region where the new fault bit is located is exhausted, selects the optimal resource from the global redundancy unit for repair based on the distance model according to the distance weight.

[0012] The beneficial effects of the present application: through the clustering algorithm to deeply mine the high-frequency fault area, and combining the comprehensive distance model fusing the physical distance and the logical distance, the precise matching and preferential allocation of the redundancy unit are realized. Among them, the preferential calling of the local redundancy unit greatly reduces the delay of cross-regional resource scheduling, and the weighted selection of the global redundancy unit effectively avoids the problem of blind allocation; the logical distance model is constructed, so that the adaptation of the redundancy unit and the fault bit is more suitable for different fault modes, such as permanent fault, intermittent fault, etc., thereby reducing the repair failure caused by improper adaptation; for the redundancy resource utilization rate, through the identification of the high-frequency fault area and the hierarchical management of the local-global redundancy pool, the problem of excessive occupation of the global redundancy resource by the local high-frequency fault is effectively avoided, and the balanced allocation of resources is realized. Attached Figure Description

[0013] The invention will now be further described with reference to the accompanying drawings.

[0014] Figure 1 This is a flowchart illustrating a local and global redundancy optimization and repair method based on data mining according to the present invention. Detailed Implementation

[0015] 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.

[0016] Please see Figure 1 As shown, this invention is a method and system for optimizing and repairing local and global redundancy based on data mining. The technical solution of this invention will be described in detail below in conjunction with the actual application scenario of NAND flash memory. In this embodiment, the NAND flash memory selected is of type 3DTLC, the storage capacity is set to 1TB, the physical partition is divided into 64 block groups, each block group is configured with 8 local redundancy units (SpareCells), and the global redundancy pool is equipped with 64 global redundancy units. The preset repair success rate threshold is 95%, and the comprehensive distance model matching degree threshold is 0.9. During the preparation phase, a fault monitoring module was first deployed to collect real-time data on NAND flash memory fault addresses, fault types (covering permanent, intermittent, and transient faults), environmental parameters (operating temperature range of -20℃ to 85℃, power supply voltage of 3.3V±0.2V), read / write frequency, and fault trigger timing. The historical fault data collection period was set to 6 months, accumulating to 1 million data entries. For algorithm tools and parameter settings, the K-Means algorithm was selected for clustering, with a preset cluster size threshold of 50 fault locations. The FP-Growth algorithm was used for association rule mining, with a support threshold of 5% and a confidence threshold of 80%. The PCA algorithm was used for dimensionality reduction, with a feature dimension threshold of 10. The analytic hierarchy process (AHP) was used to determine weights, with initial weights ω1=0.4 and ω2=0.6. Step S1, the historical fault data processing and fault bit feature extraction stage, involves cleaning the collected 1 million historical fault data entries. 12,000 invalid data entries due to voltage fluctuations and 8,000 abnormal data entries with missing time-series records are removed, leaving 980,000 valid data entries. Fault bit features are then extracted from the valid data. Fault bit coordinates are represented in the format of Block group number-page number-column number, such as Block32-Page15-Col2048, and the associated storage block is Block32. Time-series characteristics include the number of read / write operations triggered by the fault (range 1000~50000 times) and the fault duration (range 1ms~100ms). Step S2 involves cluster analysis and high-frequency fault region segmentation: In the S2.1 feature optimization process, the two core features of fault location coordinates and the memory block to which they belong are retained. The timing characteristics are transformed into auxiliary conditions. Only fault locations with a fault duration of ≥10ms are selected to exclude transient faults. The original feature dimension is 8, which is compressed to 5 dimensions after PCA dimensionality reduction.

[0017] When initializing the S2.2 clustering algorithm, the number of clusters K in the K-Means algorithm is determined to be 8 according to the elbow rule, the threshold for the number of iterations is set to 100, and the convergence error is 1e-3.

[0018] In the S2.3 clustering operation, the optimized 5-dimensional feature data is input into the initialized K-Means algorithm to generate 8 fault position clusters. Two invalid clusters with a size of less than 50 are removed, and 6 valid clusters are retained. The number of fault positions in each cluster is between 80 and 15.

[0019] In the S2.4 high-frequency fault area determination, the core attributes of each valid cluster are calculated, including center coordinates, coverage area, and fault density. Six high-frequency fault areas are determined, such as Block10-Block15, Block32-Block38, etc. After the area boundaries are determined, they are stored in the high-frequency fault area index table. The index table fields include area ID, starting block, ending block, center coordinates, and fault density. Step S3 is the construction of the distance model and the establishment of the redundant unit index table: When calculating the physical distance in S3.1, based on the physical design layout of NAND flash memory, the physical location parameters (x / y coordinates) of each redundant unit and storage unit are extracted. The Manhattan distance formula d_phys=|x1-x2|+|y1-y2| is used to calculate the physical proximity. Then, the physical distance is standardized and mapped to the [0,1] interval. For example, the physical distance between the local redundant unit of Block32 and the storage unit within the same block is standardized to 0.1~0.3, and the physical distance between redundant units across blocks is standardized to 0.6~0.9. At the same time, a penalty coefficient is set for crossing physical partitions. The penalty coefficient for crossing blocks is 0.2, and the corrected physical distance for crossing partitions is d_phys'=d_phys×(1+penalty coefficient). In the S3.2 logical distance model construction stage, the correlation between fault position features (such as fault type and timing characteristics) and fault modes (such as "high temperature environment + high frequency read / write → permanent fault") is extracted from the cleaned 980,000 historical data to form a correlation analysis dataset. The FP-Growth algorithm is used to mine association rules in the correlation analysis dataset, and strong association rules with support ≥5% and confidence ≥80% are selected, such as "fault type = permanent fault + Block = 32 → adapted redundant unit type = Type A". Based on the strong association rules, a logical distance quantification standard is defined, and a corresponding logical distance value is assigned to the matching status of the fault mode of the fault position and the adaptation mode of the redundant unit. When there is a complete match, the logical distance = 0.1; when there is a partial match, the logical distance = 0.5; and when there is no match, the logical distance = 0.9. The logical distance is then standardized to the [0,1] interval. In the S3.3 integrated distance model fusion process, the weight coefficients of physical distance and logical distance are determined by the analytic hierarchy process (AHP). Repair efficiency (correlated with physical distance) and repair stability (correlated with logical distance) are scored, ultimately determining ω1=0.35, ω2=0.65, and ω1+ω2=1. The integrated distance between the fault location and the redundant unit is calculated using the weighted summation formula D=0.35×physical distance+0.65×logical distance. For example, the integrated distance between the fault location (permanent fault) of Block32 and the local Type A redundant unit is 0.35×0.2+0.65×0.1=0.135. The matching degree of the integrated distance model is verified based on historical fault repair data. The initial model matching degree is 0.85, lower than the preset threshold of 0.9. Subsequently, the confidence threshold is adjusted to 85%, and after re-mining association rules, the matching degree improves to 0.92, meeting the preset accuracy requirements. When S3.4 establishes the redundant unit distance index table, the comprehensive distance calculation results of all redundant units after being classified into local redundant pool and global redundant pool are stored. A distance index table is constructed with the fault location area and fault mode as the index. The local redundant pool contains 64 Block groups × 8 = 512 units, and the global redundant pool contains 64 units. For example, the top 3 comprehensive distances of the local redundant units corresponding to the index "Area 3 - Permanent Fault" are [0.135, 0.152, 0.178]. Step S4 prioritizes the repair of local redundant units: S4.1 collects the fault address, fault type, storage block, real-time environmental parameters, and fault trigger timing of the new fault bit. Based on the preset fault bit feature extraction rules, it extracts the core features of the new fault bit from the collected data and forms a feature vector. The core features include the fault bit coordinates, the storage block to which it belongs, and timing characteristics. For example, the collected new fault data is fault address = Block32-Page20-Col1024, fault type = permanent fault, storage block to which it belongs = Block32, real-time environmental parameters (temperature = 65℃, voltage = 3.3V), and fault trigger timing (read / write times = 35000 times, duration = 20ms). The extracted core feature vector is [Block32, (32, 20, 1024), 35000 times, 20ms]. S4.2 Matches the memory block to which the new fault bit belongs, the coordinates of the fault bit, and the high-frequency fault area index table to determine whether the new fault bit is located in the high-frequency fault area. Then, it locates the corresponding dedicated local redundancy pool based on the physical location of the new fault bit. The local redundancy pool is pre-divided according to the physical partition of the memory. In this example, Block32 belongs to "Area 3" (high-frequency fault area), and the corresponding dedicated local redundancy pool contains 8 redundancy units. S4.3 Call the preset redundant unit distance index table, using the region and fault mode of the new fault location as the index, to query the comprehensive distance between all available redundant units in the local redundant pool and the new fault location. The comprehensive distance is the weighted sum of the physical distance and the logical distance. Sort the comprehensive distances in ascending order to obtain the top N candidate local redundant units, where N is the preset number of candidates. In this example, N=3. Using "Region 3 - Permanent Fault" as the index, query the distance index table. The comprehensive distances of the 8 units in the local redundant pool are sorted as [0.135, 0.152, 0.178, 0.210, 0.256, 0.301, 0.345, 0.390]. Filter out the top 3 candidate units. S4.4 performs working status detection, adaptation mode matching verification, and physical compatibility verification on the candidate local redundant units. The local redundant unit with the best adaptability is selected, and the fault repair process is started. The repair process includes redundancy unit activation, fault bit data migration, and address mapping update. In this example, all three candidate units are in an "idle" state and all support permanent fault repair and have no wiring conflicts. The first unit with the smallest integrated distance (distance 0.135) is selected, the redundant unit is activated, the data of fault bits Page20-Col1024 is migrated, and the original address is mapped to the redundant unit address. After S4.5 is repaired, the used local redundant units are marked as "occupied" and removed from the list of available resources. The associated records in the distance index table of redundant units are updated synchronously. If the remaining available resources in the local redundant pool are lower than the preset warning threshold, a resource warning signal is output. In this example, the local redundant pool has 7 remaining available units, which is higher than the warning threshold of 3, so no warning signal is output. Step S5 is the global redundant unit repair, simulating a local redundancy exhaustion scenario. Assume that "Region 3" continues to experience faults, and the 8 units in the local redundancy pool are gradually occupied. When the 9th new fault bit appears, it is first determined that the local redundant units in "Region 3" are exhausted (the available quantity is 0). Then, using "Region 3 - Intermittent Fault" (a new fault type) as the index, the comprehensive distance index table of the global redundant pool is queried. After sorting the comprehensive distances of the 64 global units, the top 5 candidate units (distances 0.280, 0.312, 0.335, 0.368, and 0.402) are selected. Then, status detection and adaptation verification are performed on the candidate units, and the global redundant unit with the smallest comprehensive distance (distance 0.280) is selected. The repair process is then executed, including data migration and address mapping updates. Finally, after the repair is completed, the global redundant unit is marked as "occupied," and the global redundant pool resource status and distance index table are updated. Regarding the verification of implementation effectiveness, the actual operation data of this embodiment over 6 months shows that the local redundancy repair rate reached 78%, with an average repair response time of 2.3ms, which is 60.3% higher than the traditional method (5.8ms); the global redundancy repair rate reached 22%, with an average repair response time of 4.7ms, which is 42.7% higher than the traditional random allocation method (8.2ms); the overall repair success rate was 98.5%, and the fault recurrence rate was 1.2%, which is significantly optimized compared to the traditional method (success rate 85%, recurrence rate 8.5%); the local redundancy pool resource utilization rate reached 72%, and the global redundancy pool resource utilization rate reached 45%, with no coverage insufficiency problem caused by excessive resource occupation.

[0020] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for optimizing and repairing local and global redundancy based on data mining, characterized in that, Includes the following steps: Step S1: Collect historical fault data of the memory, clean the historical fault data, and extract fault bit features; The historical fault data includes fault address, fault type, and environmental parameters; the fault bit characteristics include fault bit coordinates, the memory block to which it belongs, and timing characteristics. Step S2: Use a clustering algorithm to perform cluster analysis on the extracted fault location features and divide the high-frequency fault regions of the fault locations; Step S3: Construct a distance model between the fault location and the redundant unit. The distance model integrates the physical distance and logical distance between the fault location and the redundant unit. The physical distance is the proximity of the physical location of the redundant unit and the fault unit within the chip, and the logical distance is an abstract distance based on the fault mode, which is established through association rule mining. Step S4: When a new fault location is detected, based on the distance model, prioritize the allocation of local redundant units with closer logical and / or physical distances for repair, and update the resource occupancy status of the local redundant pool. Step S5: If the local redundant units in the area where the new fault location is located are exhausted, then the optimal resource is selected from the global redundant units for repair based on the distance model and distance weight.

2. The method for local and global redundancy optimization and repair based on data mining according to claim 1, characterized in that, Step S2 specifically includes: S2.1: From the fault location features extracted in step S1, core features that are strongly correlated with the clustering of fault areas are selected. The selection method is to retain the fault location coordinates and the storage block to which they belong, and to convert the time-series characteristics into auxiliary selection conditions. If the feature dimension is higher than the preset threshold, a dimensionality reduction algorithm is used to compress the feature dimension. S2.2: Select a clustering algorithm based on the distribution characteristics of memory fault bits, and initialize the parameters of the selected clustering algorithm; S2.3: Input the optimized feature data from step S2.1 into the initialized clustering algorithm, perform clustering operations, generate multiple fault location clusters, remove invalid clusters with a size smaller than a preset threshold, and retain valid clusters; S2.4: Calculate the core attributes for each valid cluster, determine the high-frequency fault regions based on the core attributes, clarify the region boundaries and mark and store them to form a high-frequency fault region index table.

3. The method for local and global redundancy optimization and repair based on data mining according to claim 1, characterized in that, The specific steps S3 are as follows Includes the following: S3.1: Define the physical distance calculation rules. Based on the physical design layout of the memory chip, extract the physical location parameters of redundant units and all memory units, use the Euclidean distance or Manhattan distance formula to calculate the physical proximity of redundant units and faulty units, obtain the physical distance quantification value after standardization, and combine chip wiring constraints to set a physical distance penalty coefficient for redundant units that cross physical partitions to correct the calculation results. S3.2: Construct a logical distance model through association rule mining; S3.3: Integrate physical distance and logical distance to construct a comprehensive distance model; S3.4: Establish a distance index table for redundant units, store the comprehensive distance calculation results of all redundant units after classifying them into local redundancy pools and global redundancy pools, and construct a distance index table indexed by the region to which the fault location belongs and the fault mode.

4. The method for local and global redundancy optimization and repair based on data mining according to claim 3, characterized in that, Step S3.2 specifically includes the following: S3.2.1: Based on the historical fault data cleaned in step S1, extract the correlation between fault location features and fault modes to form a correlation analysis dataset; S3.2.2: Use the Apriori algorithm or FP-Growth algorithm to mine association rules in the association analysis dataset and filter out strong association rules with confidence and support higher than the preset threshold; S3.2.3: Define a logical distance quantification standard based on strong correlation rules, assign corresponding logical distance values ​​to the matching status of the fault mode of the fault position and the adaptation mode of the redundant unit, and then standardize the logical distance to a preset range.

5. The method for local and global redundancy optimization and repair based on data mining according to claim 3, characterized in that, Step S3.3 specifically includes the following: S3.3.1: Determine the weight coefficients of physical distance and logical distance through the analytic hierarchy process or machine learning modeling based on historical repair success rates, and balance repair efficiency and repair stability in weight allocation; S3.3.2: The weighted summation formula is used to calculate the comprehensive distance between the fault location and the redundant unit. The weighted summation formula is D=ω1×physical distance+ω2×logical distance, where ω1 is the physical distance weighting coefficient, ω2 is the logical distance weighting coefficient, and ω1+ω2=1. S3.3.3: Verify the matching degree of the comprehensive distance model based on historical fault repair data. If the matching degree is lower than the preset threshold, adjust the weight coefficient or the threshold parameter of association rule mining, and iterate and optimize until the model meets the preset accuracy requirements.

6. The method for local and global redundancy optimization and repair based on data mining according to claim 2, characterized in that, Step S4 specifically includes: S4.1: Collect the fault address, fault type, storage block to which it belongs, real-time environmental parameters and fault trigger timing of the new fault location. Based on the preset fault location feature extraction rules, extract the core features of the new fault location from the collected data and form a feature vector. The core features include the fault location coordinates, the storage block to which it belongs and timing characteristics. S4.2: Match the storage block to which the new fault bit belongs, the coordinates of the fault bit, and the high-frequency fault area index table to determine whether the new fault bit is located in the high-frequency fault area. Then, locate the corresponding dedicated local redundancy pool according to the physical location of the new fault bit. The local redundancy pool is pre-divided according to the physical partition of the memory. S4.3: Call the preset redundant unit distance index table, and use the region and fault mode of the new fault position as the index to query the comprehensive distance between all available redundant units in the local redundant pool and the new fault position. The comprehensive distance is the weighted sum of the physical distance and the logical distance. Sort and filter the top N candidate local redundant units in ascending order of comprehensive distance, where N is the preset number of candidates. S4.4: Perform working status detection, adaptation mode matching verification and physical compatibility verification on the candidate local redundant units, select the local redundant unit with the best adaptability, and start the fault repair process. The repair process includes redundancy unit activation, fault bit data migration and address mapping update. S4.5: After the repair is completed, the used local redundant units are marked as "occupied" and removed from the list of available resources. The associated records in the distance index table of the redundant units are updated synchronously. If the remaining available resources in the local redundant pool are lower than the preset warning threshold, a resource warning signal is output.

7. A local and global redundancy optimization and repair system based on data mining, characterized in that, include: Data acquisition module: Collects historical fault data of the memory, cleans the historical fault data, and extracts fault bit features; The historical fault data includes fault address, fault type, and environmental parameters; the fault bit characteristics include fault bit coordinates, the memory block to which it belongs, and timing characteristics. Clustering analysis module: The clustering algorithm is used to perform clustering analysis on the extracted fault location features to divide the high-frequency fault regions of the fault locations; Model building module: Constructs a distance model between fault locations and redundant units. The distance model integrates the physical distance and logical distance between the fault locations and redundant units. The physical distance is the proximity of the physical locations of the redundant units and fault units within the chip, and the logical distance is an abstract distance based on fault modes, established through association rule mining. First repair module: When a new fault location is detected, based on the distance model, local redundant units with closer logical and / or physical distances are preferentially allocated for repair, and the resource occupancy status of the local redundant pool is updated. Second repair module: If the local redundant units in the area where the new fault location is located are exhausted, the optimal resource is selected from the global redundant units for repair based on the distance model and distance weight.

Images