A method for evaluating the quality of wood-rotting edible fungus spores

Abstract

The present application relates to edible fungus strain quality evaluation and culture process time series data processing technical field, specifically to a kind of wood-rotting edible fungus strain quality evaluation method.Method obtains multivariate time series data in the process of strain culture, carries out trend maintenance preprocessing and generates change sequence;According to the synergistic relationship of multiple change sequences in adjacent time window, determine the germination start node and stable expansion node, map the absolute culture time to biological process coordinates with the node as anchor point, complete growth stage alignment;From the stage after alignment, extract transition topological features, input time series pattern recognition model with stage order constraint and growth trend constraint, obtain germination capacity prediction value and strain activity evaluation value, and generate strain quality evaluation result combined with stage state path consistency.The present application can reduce the stage dislocation caused by the difference in germination time, distinguish early germination state and continuous growth state, so that the evaluation result reflects strain germination capacity and continuous activity at the same time.

Description

Technical Field

[0001] This invention relates to the field of edible fungi strain quality evaluation and time-series data processing technology, specifically a method for evaluating the quality of wood-rotting edible fungi strains. Background Technology

[0002] The quality evaluation of wood-rotting edible fungi strains typically focuses on the mycelial growth process in petri dishes, test tubes, or culture bags. This involves periodically recording indicators such as mycelial coverage area, colony diameter, edge advancement distance, texture density, and growth uniformity to assess strain vigor and germination ability. Current conventional methods often involve collecting images or manually measuring data at uniform incubation times, smoothing the original sequences, calculating the difference between adjacent time points, average growth rate, cumulative growth, or time required to reach a preset coverage area, and outputting a quality grade based on empirical thresholds, statistical discrimination models, or time-series classification models. Some methods normalize multiple indicators and perform weighted summation, or input the complete incubation sequence into a time-series model to learn the numerical differences between different quality grades. The common basis of these methods is using absolute incubation time as the sample alignment scale, treating different strains at the same time point after inoculation as being in the same growth stage, and then comparing their numerical values ​​or trends. To reduce the impact of data acquisition fluctuations, fixed window filtering, outlier removal, and missing value imputation are usually used. However, the processed data still retains the original absolute time coordinates, and the evaluation results mainly depend on the local performance at the preset time or the average performance over the entire process.

[0003] Directly aligning different fungal samples based on absolute incubation time cannot eliminate the misalignment of biological processes caused by differences in the timing of germination initiation. Wood-rotting edible fungi typically undergo continuous stages after inoculation, including adaptation, germination initiation, mycelial expansion, and stable growth. Different samples enter each stage at different times; at the same incubation time, some samples may still be in the adaptation stage, while others may have entered the expansion stage. Directly comparing coverage area, growth rate, or texture changes at that time will introduce differences in stage position into quality differences, leading to underestimation of species that germinate later but subsequently expand stably, or overestimation of species that germinate earlier but lack sustained growth capacity. Fixed-window smoothing and average growth rate calculations can only reduce local noise and cannot identify the true transition positions from the adaptation stage to the initiation stage, or from the initiation stage to the expansion stage. Even using complete sequences, ordinary time-series classification models may establish superficial associations between incubation time points and quality labels. Therefore, existing evaluation results cannot simultaneously reflect germination initiation capacity and sustained growth vigor, primarily due to the lack of a stage alignment mechanism based on biological process nodes. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating the quality of wood-rotting edible fungi strains, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for evaluating the quality of wood-rotting edible fungi strains includes: acquiring multivariate time-series data of the cultivation process of wood-rotting edible fungi strains, performing trend-preserving preprocessing on the time-series data of the cultivation process, and generating a change sequence; Based on the collaborative change relationship of multiple change sequences within adjacent time windows, the germination initiation node and stable expansion node are determined. Using the germination initiation node and the stable expansion node as anchor points, the absolute culture time is mapped to biological process coordinates and the growth stages are aligned. Extract the stage transition topological features from each aligned growth stage; The stage transition topological features are input into a temporal pattern recognition model constructed based on a 3-layer causal convolutional network, which includes intra-stage trajectory coding branches and inter-stage transition coding branches, and has stage sequence constraints and growth trend constraints. This model obtains germination ability prediction values ​​and strain vigor evaluation values, and generates strain quality evaluation results based on the consistency of stage state paths.

[0006] Preferably, the trend preservation preprocessing includes: constructing an adaptive smoothing window according to the local monotonic intervals of each variable, determining the smoothing intensity based on the directional consistency of adjacent sampling points within the window, suppressing isolated fluctuations where the directional consistency does not meet the continuous change condition, and retaining the original turning point for continuous unidirectional change intervals. Based on the preprocessed time-series data of the cultivation process, first-order change sequences, continuous change sequences, and local stability sequences are formed respectively, and the three types of sequences are combined according to the same time index to form a joint time-series representation for transition identification.

[0007] Preferably, determining the germination initiation node and the stable expansion node includes: calculating the change direction, duration, and relative increase of each change sequence within a sliding time window, and constructing a directed weighted cooperative change subgraph; the cooperative change subgraph uses each culture variable as a node, and the node attributes include relative increase, duration, local stability, and data integrity; when two culture variables have the same change direction and their duration change intervals overlap, a directed edge is established between the corresponding nodes, and the edge attributes include the overlap time length, the consistency of change direction, and the change initiation order; and a transition candidate interval is formed based on the node retention rate and edge continuation rate of the cooperative change subgraph between consecutive windows. Based on the node preservation relationship and edge connection continuation relationship of the cooperative change subgraph between consecutive windows, a transition candidate interval is formed; Perform a state separability test on the candidate transition intervals. The first candidate transition interval that passes the test is determined as the germination initiation node, and the candidate transition intervals that maintain the growth state for the first time are determined as the stable expansion nodes.

[0008] Preferably, the growth stage alignment includes: dividing each bacterial sample into an adaptation stage, an initiation stage, an expansion stage, and a stable stage based on the germination initiation node and the stable expansion node; For each stage, a segmented monotonic mapping relationship that maintains the time order is established, and the corresponding stages of different bacterial samples are compressed or extended to a unified stage coordinate range. At the stage boundary, a continuous transition weight determined by the node confidence is set, and the trajectory features on both sides of the boundary are fused according to the continuous transition weight to form an aligned temporal trajectory with a unified biological process scale.

[0009] Preferably, the construction of the joint time series representation includes: establishing extreme value preservation markers for local extreme points in the first-order change sequence, establishing continuous segment markers for continuous segments in the same direction in the continuous change sequence, and establishing stability boundary markers for variance abrupt change locations in the local stability sequence. The three types of markers are encoded into multi-layer trajectory segments according to their temporal adjacency. The multi-layer trajectory segments are then sorted using the synchronous occurrence order of different variables to obtain trajectory feature extraction results that simultaneously include turning points, continuous states, and stable boundaries.

[0010] Preferably, the separability test of the preceding and following states includes: extracting multivariate joint distribution features on the front and back sides of the transition candidate interval respectively, and constructing state prototypes on both sides of the candidate interval; Calculate the attribution sequence of each time point relative to the two state prototypes, and detect whether the attribution sequence forms a unidirectional continuous transition from the previous state to the next state; Node confidence is generated based on the node retention rate, edge continuation rate, and unidirectional continuous transition length of the cooperative change subgraph, and only the transition candidate intervals that satisfy the cross-window continuity condition are retained.

[0011] Preferably, when a bacterial sample is missing either the germination initiation node or the stable expansion node, a reference trajectory matching the preceding trajectory of the missing node is selected from the bacterial samples in the same batch that have completed node identification, and the candidate position of the missing node is inferred based on the local morphological correspondence between the preceding trajectories. The estimated confidence of the candidate positions and the node confidence of the identified nodes are used together to determine the boundary weights of the piecewise monotonic mapping, and the original time proportions are retained for stages where the estimated confidence does not meet the continuous mapping condition.

[0012] Preferably, the temporal pattern recognition model is based on a 3-layer causal convolutional network coding structure, including intra-stage trajectory coding branches and inter-stage transition coding branches; The intra-stage trajectory coding branch extracts the growth retention, fluctuation clustering and terminal decay trend of each stage according to the unified stage coordinates; the inter-stage transition coding branch extracts the duration of the adaptation stage, the steepness of the initiation transition and the degree of multivariate synchronization according to the stage sequence. A stage boundary confidence score is generated based on the continuity of the trajectories on both sides of the adjacent stage boundary. The stage boundary confidence score is used as the gating signal for gating fusion. The fusion weight of the inter-stage transition coding branch is determined by the stage boundary confidence score, and the fusion weight of the trajectory coding branch within the stage is determined by the weight complementary to the stage boundary confidence score. Gating fusion is performed on the outputs of the two coding branches according to the fusion weight.

[0013] Preferably, the consistency determination of the stage state path includes: representing the adaptation stage, startup stage, expansion stage and stabilization stage as ordered state nodes respectively, and constructing an ordered state diagram that only allows the state to remain or to transition sequentially along the adaptation stage, startup stage, expansion stage and stabilization stage; A state transition vector is generated based on the encoding results of adjacent stages. The state transition vector is then mapped to the ordered state diagram. Conflict segments corresponding to premature decay, repeated startup, and cross-stage reverse transition are identified. Path conflict representations are generated based on the stage, duration interval, and stage boundary confidence of the conflict segments.

[0014] Preferably, the path conflict characterization, the germination ability prediction value, and the strain viability evaluation value are input into the consistency calibration layer; The consistency calibration layer encodes conflict fragments before the initiation phase as germination calibration components that only apply to the germination capability dimension, encodes conflict fragments in the expansion and stabilization phases as continuous vitality calibration components that only apply to the continuous vitality dimension, and encodes cross-phase reverse transfers as joint constraint components that apply to both dimensions simultaneously; the values ​​of the three calibration components range from -1 to 0, and their values ​​are positively correlated with the duration of the conflict fragment. The three calibration components are combined according to the node confidence and stage boundary confidence corresponding to each conflict segment, and the combined calibration components are jointly encoded with the output of the time-series pattern recognition model to form the strain quality evaluation result.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Based on the synergistic changes in time-series data of multivariate culture processes, germination initiation nodes and stable expansion nodes are identified. Using these nodes as anchor points, the absolute culture time of different bacterial strains is converted into a unified biological process coordinate system, allowing comparisons of the adaptation, initiation, expansion, and stable phases within their respective intervals. Thus, the evaluation model no longer directly considers identical time points after inoculation as identical growth states, but rather determines the sample's stage based on the actual state transition location. Intra-stage trajectories and inter-stage transition characteristics are identified under a unified coordinate system, allowing the evaluation to incorporate germination start time, initiation transition intensity, expansion maintenance state, and stable phase decay trend. Outputs that do not conform to a continuous growth process are limited by stage sequence constraints and growth trend constraints. This processing eliminates the confusion caused by differences in germination time in numerical comparisons, creating different evaluation paths for samples that germinate later but expand stably and those that germinate earlier but subsequently decay. This ensures that the bacterial strain quality results simultaneously correspond to germination capacity and sustained viability, resolving the stage misalignment problem caused by absolute time alignment.

[0016] 2. Trend Preservation Preprocessing: The smoothing intensity is adjusted according to local monotonic intervals to suppress isolated fluctuations while preserving true turning points. A joint temporal representation is formed by first-order change sequences, continuous change sequences, and locally stable sequences, providing a continuous and mutually verified data foundation for transition node identification. The Cooperative Change Subgraph, combining node preservation relationships, edge connection continuation relationships, and the separability of preceding and following states, can eliminate false transitions caused by short-term fluctuations in a single indicator. Node confidence and stage boundary confidence are used to control the combination of stage mapping, boundary feature fusion, and dual-branch encoding results, reducing the transmission of uncertain boundaries to subsequent evaluations. For samples with missing nodes, reference trajectory matching and preservation of the original time ratio jointly limit the inference range, avoiding trajectory distortion caused by forced alignment. The Ordered State Graph further identifies conflict segments formed by repeated initiation, premature decay, and cross-stage reverse transitions, and uses conflict information for calibration of germination ability and sustained vitality, ensuring the final result retains a complete stage state path and reducing the interference of local high growth or single-stage anomalies on overall quality judgment. Attached Figure Description

[0017] Figure 1 A flowchart for the overall quality evaluation of wood-rotting edible fungi strains; Figure 2 Plot of original observations and trend-preserving values; Figure 3 To standardize the coordinate diagram of biological processes; Figure 4 This is a graph showing the state attribution probability and node confidence. Detailed Implementation

[0018] In one embodiment, please refer to Figure 1A method for evaluating the quality of wood-rotting edible fungi strains includes: acquiring multivariate time-series data of the cultivation process of wood-rotting edible fungi strains; performing trend-preserving preprocessing on the time-series data to generate change sequences; determining germination initiation nodes and stable expansion nodes based on the collaborative change relationships of multiple change sequences within adjacent time windows; mapping absolute cultivation time to biological process coordinates and aligning growth stages using the germination initiation nodes and stable expansion nodes as anchor points; extracting stage transition topological features from each aligned growth stage; inputting the stage transition topological features into a time-series pattern recognition model constructed based on a 3-layer causal convolutional network, containing intra-stage trajectory coding branches and inter-stage transition coding branches, and having stage sequence constraints and growth trend constraints, to obtain germination capacity prediction values ​​and strain vigor evaluation values, and generating strain quality evaluation results based on the consistency of stage state paths.

[0019] In this embodiment, the time-series data of the multivariate culture process are associated according to the same strain, the same culture batch, and the same collection time. The data variables include at least three of the following obtained from the culture process: hyphal coverage, hyphal edge advancement distance, growth per unit time, hyphal distribution uniformity, hyphal texture density, and growth region continuity. The hyphal coverage is used to indicate the extent to which the hyphae actually occupy the culture area; the hyphal edge advancement distance is used to indicate the spatial expansion state of the growth front; the growth per unit time is used to indicate the increment between adjacent collection times; the hyphal distribution uniformity is used to distinguish between local concentrated growth and overall synchronous expansion; the hyphal texture density is used to reflect the tissue state after hyphal formation; and the growth region continuity is used to identify whether discrete patches have connected to form a stable growth region. Each variable is registered using timestamps and sample identifiers. When timestamps are inconsistent, the nearest time matching is performed based on the minimum common collection interval, and data exceeding the allowed matching interval is retained as missing data and not directly interpolated across stages.

[0020] Specifically, the preprocessing of time-series data during the cultivation process is constrained by maintaining the state transition position. For each variable, the local change direction and continuous interval are detected separately. Smoothing is performed in intervals where the local change direction remains consistent, and the smoothing degree is reduced in intervals where the change direction changes or adjacent increment signs continuously conflict, so as to avoid spreading the true transition before and after germination to multiple collection times. For isolated outliers, correction is made based on the deviation relationship between the outlier and the continuous trajectory before and after. If the change direction before and after the outlier is consistent and the outlier deviates from the local trajectory, the interpolation result of the trajectory before and after is used to replace the outlier. If the outlier occurs at a time when multiple variables change simultaneously, the outlier is retained and the corresponding time is marked as the transition time to be determined. For continuous missing segments, trend constraint interpolation is performed only when the two sides of the missing segment belong to the same local monotonic interval. When the missing segment crosses the candidate transition interval, no definite value is generated. Instead, upper and lower bounds of the interval are generated for subsequent node confidence calculation.

[0021] In one specific implementation, the trend-smoothing result is calculated according to the following formula: ; ; in, This represents the original value of the i-th culture variable at the t-th data collection time. This represents the corresponding smoothing value, and q represents the adjacent position relative to the current time. This represents the window radius determined based on the direction of local change. This represents the weighting coefficients for adjacent data. This indicates the proportion of adjacent increments with the same sign within the current window; for example, when consecutive increments within a window have the same direction, Take values ​​close to 1 and allow a wider window to participate in the calculation. When there is a continuous transition from negative to positive change within the window, By reducing and decreasing the weight of data that is far from the current time, a continuous trajectory is obtained in the stable interval, while the inflection interval retains the germination start position.

[0022] Please refer to Figure 2 , Figure 2 The horizontal axis represents the sequential data collection from time 1 to time 5, and the vertical axis represents the observed values ​​and trend maintenance values ​​of each culture variable. In the legend, coverage, edge, growth, uniformity, density, and continuity correspond to hyphal coverage, edge advancement, growth per unit time, distribution uniformity, texture density, and regional continuity, respectively. The original corresponding formula symbols are also provided. Trend corresponds to trend persistence result .

[0023] from Figure 2It is evident that the trend curves of most variables, while maintaining an upward trend, weaken local spikes, indicating that the weights... and window radius It is primarily used to suppress sampling fluctuations, rather than altering the true growth inflection point. The coverage and growth correlation curves rise more significantly in the latter part, indicating... The proportion of unidirectional changes reflected gradually increases, and subsequent node identification can continue to generate first-order change sequences, continuous change sequences, and locally stable sequences based on this stable change.

[0024] After completing the trend-preserving preprocessing, a first-order change sequence, a persistent change sequence, and a local stability sequence are generated for each variable. The first-order change sequence consists of the difference between the current smoothed value and the previous smoothed value, used to record the direction of change and the instantaneous change amplitude. The persistent change sequence accumulates the duration based on whether adjacent first-order change values ​​have the same sign, and recounts when the change direction changes, used to distinguish between single-moment fluctuations and continuous growth trends. The local stability sequence consists of the dispersion of each smoothed value within the current window relative to the local median trajectory, used to record the position where the trajectory transitions from a fluctuating state to a stable state. The three types of sequences are combined using the same time index, and a scaling transformation based on the steady-state interval within the batch is performed on the value ranges of different variables, so that the scaling transformation only eliminates the difference in numerical dimensions without changing the turning point position, duration, and relative order of change of the variables on the time axis.

[0025] In a preferred embodiment, the trend preservation preprocessing includes: constructing an adaptive smoothing window based on the local monotonic intervals of each variable; determining the smoothing intensity based on the directional consistency of adjacent sampling points within the window; suppressing isolated fluctuations where the directional consistency does not meet the continuous change condition; and retaining the original turning points for continuous unidirectional change intervals; forming first-order change sequences, continuous change sequences, and locally stable sequences based on the preprocessed time-series data of the cultivation process; and combining the three types of sequences according to the same time index into a joint time-series representation for transition identification; the local monotonic interval is determined by detecting the sign of continuous increments and the cumulative deviation; when there is only a single reverse increment in the continuous increments and the reverse increment does not cause a change in the direction of the subsequent trajectory, the reverse increment is identified as an isolated fluctuation; when a continuous reverse trajectory is formed after the reverse increment, the local monotonic interval is broken at the direction change point, and the boundary of the adaptive smoothing window must not cross the direction change point.

[0026] Furthermore, each acquisition moment in the joint time series representation includes the first-order change value, duration, local dispersion, and corresponding data integrity marker for each culture variable. The data integrity marker is used to distinguish between measured values, interpolated values ​​within the same interval, and uncertain values ​​across intervals. When calculating the collaborative change relationship of multiple variables, measured values ​​participate according to their complete weight, interpolated values ​​within the same local monotonic interval participate in descending order of interpolation segment length, and uncertain values ​​across intervals do not participate in the change direction voting but participate in confidence deduction. Thus, node identification does not rely on a single complete sequence but forms a joint judgment based on the change direction, duration, and data reliability of each variable in the same time neighborhood, avoiding the direct interruption of the entire culture trajectory due to the missing acquisition of individual variables.

[0027] The degree of coordinated change of multiple change sequences within adjacent time windows is calculated using the following formula: ; in, This represents the collaborative change value at the t-th data collection time, where M represents the number of culture variables involved in the determination. This represents the smoothed difference between adjacent values ​​of the i-th variable. This indicates the main direction of change within the current window, determined by the majority of variables. This is an instruction operation that takes a value of 1 when the condition is true and a value of 0 when the condition is false. This represents the persistence of the i-th variable in the current direction. This represents the standardized variation of the i-th variable relative to its stable fluctuation range; for example, when hyphal coverage, edge advancement distance, and texture density all change positively over multiple consecutive time points, the indicator operations of the three variables are set to 1 and form a higher value based on the persistence. If only the mycelial coverage area increases suddenly at a single moment while other variables do not change synchronously, then only the corresponding variable participates in the summation, and the current moment is not directly identified as a biological stage transition.

[0028] In one embodiment, determining the germination initiation node and the stable expansion node includes: calculating the change direction, duration, and relative increase of each change sequence within a sliding time window, and constructing a directed weighted cooperative change subgraph; the cooperative change subgraph uses each culture variable as a node, and the node attributes include relative increase, duration, local stability, and data integrity; when two culture variables have the same change direction and their duration change intervals overlap, a directed edge is established between the corresponding nodes, and the edge attributes include the overlap time length, the consistency of change direction, and the change initiation order; a transition candidate interval is formed based on the node retention rate and edge continuation rate of the cooperative change subgraph between consecutive windows; a transition candidate interval is formed based on the node retention relationship and edge connection continuation relationship of the cooperative change subgraph between consecutive windows; a state separability test is performed on the transition candidate intervals, and the first transition candidate interval that passes the test is determined as the germination initiation node, and the transition candidate interval whose subsequent growth retention state first reaches stability is determined as the stable expansion node.

[0029] The cooperative change subgraph uses the cultivated variable as the graph node. The condition for connecting two variables is that they have the same main change direction within the same window and their duration overlaps over time. Node attributes include relative increase, duration, local stability, and data integrity. Edge attributes include overlap duration, consistency of change direction, and change initiation order. Adjacent sliding windows establish node correspondences according to variable identifiers. When the same variable maintains the same change direction in consecutive windows, a node maintenance relationship is formed. When the connection between two variables persists in consecutive windows, an edge connection continuation relationship is formed. If the subgraphs in multiple consecutive windows maintain a common set of nodes and the edge connections do not break in an unordered manner, the corresponding time ranges are merged into transition candidate intervals. The starting point of the candidate interval is the moment when the common nodes first form a stable connection, and the ending point of the candidate interval is the moment when the main change direction begins to stabilize or the cooperative connection begins to weaken.

[0030] The determination of the germination initiation node is not based on the first increase of a single variable, but requires that the trajectory before the candidate transition interval is in a low-change or unstable adaptive state, that multiple variables within the candidate transition interval form a continuous positive synergistic change, and that at least one of the following—hyphae coverage, edge advancement distance, or growth area continuity—continues to maintain a positive change after the candidate transition interval. The determination of the stable expansion node uses the candidate interval after the germination initiation node as the search range, requiring that the growth trajectory before the candidate interval still has a large fluctuation in the initiation stage, that the growth per unit time after the candidate interval enters a continuous range, and that the evenness of hyphal distribution and the continuity of the growth area no longer show continuous changes opposite to the expansion direction. When there are multiple candidate intervals for the same culture trajectory, the node is selected according to the separability of the preceding and following states, the continuity of synergistic changes, and the stage sequence relationship, rather than directly selecting based on the local increase of a single candidate interval.

[0031] In one embodiment, the growth stage alignment includes: dividing each bacterial sample into an adaptation stage, a start-up stage, an expansion stage, and a stable stage based on germination initiation nodes and stable expansion nodes; establishing a segmented monotonic mapping relationship that maintains the temporal order for each stage, compressing or extending the corresponding stages of different bacterial samples to a unified stage coordinate range; setting continuous transition weights determined by node confidence at the stage boundaries, and fusing trajectory features on both sides of the boundaries based on the continuous transition weights to form an aligned temporal trajectory with a unified biological process scale.

[0032] The time mapping for each stage is performed according to the following formula: ; in, This represents the position of the t-th data collection time in the unified biological process coordinate system. and These represent the starting and ending boundaries of the j-th stage, respectively. and These represent the start and end positions of the j-th stage in a unified coordinate system, where j corresponds sequentially to the adaptation, initiation, expansion, and stabilization stages. For example, the initiation stage of one species lasts for 8 sampling intervals, while the initiation stage of another species lasts for 13 sampling intervals; both are mapped to the same coordinate system. During the mapping process, the sequence of collection times and the direction of local changes are maintained, so that subsequent comparisons are made for the same biological stage without relying on the absolute culture time.

[0033] Please refer to Figure 3 The horizontal axis represents continuous sampling points in chronological order during the sample culture process, and the vertical axis represents the coordinates of a uniform biological process. The solid line is obtained by mapping from the absolute culture time. The dashed line and dotted line represent the stage boundary confidence and transition weight, respectively, which are used to explain whether the boundary neighborhood is suitable for direct compression or expansion.

[0034] Figure 3 middle The values ​​gradually approach 1 from low values ​​over time, indicating the segmentation boundaries. and Once determined, each collection point is mapped to a location defined by [the map / database name]. and Within the defined stage intervals, the boundary confidence levels remained generally high, indicating that the germination initiation nodes and stable expansion nodes support the stage division. The transition weight curves were relatively smooth, suggesting that no significant breaks occurred at the stage boundaries. Subsequently, the adaptation, initiation, expansion, and stabilization stages of different bacterial species can be compared at the same biological process scale.

[0035] The continuous transition at the stage boundaries is achieved through a two-sided trajectory fusion approach. Boundary neighborhoods are set on both sides of the germination initiation node and the stable expansion node. When the node confidence is high, the boundary neighborhood is reduced and the trajectory before and after the node is clearly segmented. When the node confidence is low, the boundary neighborhood is expanded and the features on both sides are weighted according to the confidence level. The fusion of the end of the adaptive stage and the beginning of the initiation stage only considers the local growth trend and fluctuation state, without averaging the duration. The fusion of the end of the initiation stage and the beginning of the expansion stage considers both the growth rate and stability changes. The aligned trajectory after fusion retains the relative time order within each stage and forms continuous feature values ​​at the stage boundaries to avoid numerical jumps caused by segmented mapping, which may be mistaken by the temporal pattern recognition model as new growth events.

[0036] When extracting the topological features of stage transitions from each aligned growth stage, stages are used as ordered nodes, and state transitions between adjacent stages are used as directed connections. Adaptation stage nodes include stage duration, local fluctuation aggregation degree, and terminal change direction; initiation stage nodes include germination initiation steepness, multivariate synchronization degree, and change persistence; expansion stage nodes include growth preservation degree, spatial expansion continuity, and number of local fallbacks; and stable stage nodes include terminal decay trend, trajectory dispersion, and stable interval length. Directed connections record the boundary confidence, change direction preservation relationship, variable synchronization order, and stage transition integrity between adjacent stages. Stage transition integrity is determined based on whether the terminal state of the previous stage continuously enters the starting state of the next stage. Trajectories that have not completed the expected transition are retained as conflict connections, and the original state path is not changed by deleting abnormal segments.

[0037] In a preferred embodiment, the construction of the joint temporal representation includes: establishing extremum preservation markers for local extrema in a first-order change sequence, establishing continuous segment markers for continuous segments in the same direction in a continuous change sequence, and establishing stability boundary markers for variance abrupt change locations in a local stability sequence; encoding the three types of markers into multi-layer trajectory segments according to temporal adjacency, and sorting the multi-layer trajectory segments using the synchronous occurrence order between different variables to obtain trajectory feature extraction results that simultaneously contain turning points, continuous states, and stability boundaries.

[0038] Specifically, the extreme value retention marker is determined by the positional drift of the local extreme point in multiple adjacent windows. If the extreme point remains in the same temporal neighborhood after the window moves, the extreme point is retained. If the extreme point drifts disorderly with the window movement, it is marked as a local fluctuation. The continuous segment marker consists of the start and end positions of continuous unidirectional increments, the cumulative change, and the interruption method. The continuous segment is not interrupted when the original direction is restored after a single reverse increment. The original continuous segment ends when multiple continuous reverse increments form a new directional state. The stability boundary marker consists of the position where the local dispersion changes from a high fluctuation state to a low fluctuation state or from a low fluctuation state to a continuously decaying state. The three types of markers are connected according to the temporal adjacency relationship within the same variable to form a variable trajectory layer, and then connected according to the order of appearance between different variables to form a cross-variable synchronization layer, so that the order of germination and initiation, the continuous state of continuous expansion, and the boundary position of the stable stage are simultaneously entered into subsequent node identification and model encoding.

[0039] In one embodiment, the separability test of the preceding and following states includes: extracting multivariate joint distribution features from the front and back sides of the candidate transition interval, respectively, and constructing state prototypes on both sides of the candidate interval; calculating the attribution sequence of each time point relative to the two state prototypes, and detecting whether the attribution sequence forms a unidirectional continuous transition from the preceding state to the following state; generating node confidence based on the node retention rate, edge continuation rate, and unidirectional continuous transition length of the cooperative change subgraph, and retaining only the candidate transition intervals whose node confidence satisfies the cross-window continuity condition.

[0040] The preceding state prototype is composed of the median vector of the multivariate joint features in a continuous stable window before the starting point of the candidate interval, and the following state prototype is composed of the median vector of the multivariate joint features in a continuous stable window after the ending point of the candidate interval. The degree of belonging to either state prototype is calculated at each time step within the candidate interval, as follows: ; in, This indicates the degree to which the state at time t belongs to the prototype of the state at time k. This represents a joint eigenvector composed of the direction of change, duration, relative increase, and local stability of multiple variables. This indicates the prototype of the preceding state or the prototype of the following state. This represents the discrete scale within the corresponding state prototype. Indicates the index of the two state prototypes. Indicates the first A state prototype State prototype Internal discrete scale; for example, the front of the candidate interval When approaching the prototype of the previous state, the degree of belonging to the previous state is relatively high. As the cultivation trajectory enters the germination and initiation process, the degree of belonging to the subsequent state continuously increases and remains. The belonging sequence forms a unidirectional transition. If the two degrees of belonging change alternately at multiple times, the candidate interval is identified as a fluctuation interval and not as a definite node.

[0041] Please refer to Figure 4 The horizontal axis represents the continuous collection order within the candidate interval, from time 1 to time 5, and the vertical axis represents the state assignment probability and node confidence. Solid lines represent the degree of state assignment for the preceding state, dashed lines represent the degree of state assignment for the following state, and dotted lines represent node confidence. From joint eigenvectors relative to state prototype Distance and Discrete Scale To be determined jointly.

[0042] from Figure 4 It can be seen that the pre-attribution curve gradually decreases from a high value, while the post-attribution curve gradually increases from a low value, forming a clear transition near the intermediate collection point. This indicates that there is a unidirectional and continuous transition from the previous state to the subsequent state within the candidate interval. The node confidence remains at a high level, indicating that this transition is not a random result caused by short-term fluctuations of a single variable, but rather is consistent with the node retention rate, edge continuation rate, and state prototype separability in the collaborative change subgraph. This can be used as a basis for determining whether a germination initiation node or a stable expansion node is a valid node.

[0043] The node retention rate is determined by the ratio of the number of common variable nodes to the total number of variable nodes in the continuous window. The edge continuation rate is determined by the ratio of the number of common edges in the continuous window to the total number of edges in the candidate interval. The unidirectional continuous transition length is determined by the number of consecutive moments in which the degree of belonging to the subsequent state is continuously greater than that of the previous state. The node confidence is jointly encoded by three results and data integrity. When the data integrity is low, the upper bound of the node confidence is limited. When the discrete scale of the prototypes of the preceding and following states is too large, the sampling length on both sides of the candidate interval is increased. If a stable prototype still cannot be formed, the candidate interval is retained but the node position is uncertain. The germination start node is the center position of the first candidate interval that passes the separability test and satisfies the sequential transition from the adaptation stage to the start stage. The stable expansion node is the center position of the first candidate interval that passes the separability test after the germination start node and whose subsequent state is continuously maintained.

[0044] In one embodiment, when a bacterial sample is missing either a germination initiation node or a stable expansion node, a reference trajectory matching the preceding trajectory of the missing node is selected from the bacterial samples in the same batch that have completed node identification. The candidate position of the missing node is inferred based on the local morphological correspondence between the preceding trajectories. The inferred confidence of the candidate position and the node confidence of the identified node are used together to determine the boundary weight of the piecewise monotonic mapping, and the original time proportion is retained for stages where the inferred confidence does not meet the continuous mapping condition.

[0045] The selection of reference trajectories uses the observed trajectories before the missing node as the matching object. Normalized morphological sequences are extracted for hyphal coverage, edge advancement distance, growth per unit time, and local stability, and candidate reference samples are determined by local path matching that maintains the time order. Candidate reference samples must belong to the same culture batch or have comparable data acquisition semantics, and the trajectory change direction, continuous segment order, and stability boundary order before the identified node must be consistent with the sample to be evaluated. When estimating the missing node, the relative position of the corresponding node in the reference trajectory with respect to the previous matching endpoint is mapped to the sample to be evaluated, and a candidate position set is formed by multiple candidate reference samples. A higher estimation confidence is formed when the candidate position set is concentrated and the previous matching is continuous, and the estimation confidence is reduced when the candidate positions are scattered or the previous matching only covers a short interval.

[0046] In cases where a germination initiation node is missing but a stable expansion node has been identified, the presumed position must be earlier than the stable expansion node and satisfy the minimum structural integrity of the initiation phase. The minimum structural integrity includes at least one interval of multivariate unidirectional change and one interval of transition from high volatility to continuous growth. In cases where a germination initiation node has been identified but a stable expansion node is missing, the presumed position must be later than the germination initiation node and be within the time range where growth volatility begins to converge. When the presumed confidence does not meet the continuous mapping condition, no forced compression or expansion is performed on the stage corresponding to the missing node. Instead, the stage coordinates are determined according to the proportion of the original duration to the complete culture duration, and the stage is marked as a low boundary confidence state, so that subsequent models can reduce the participation of boundary uncertainty information through gating fusion.

[0047] In one embodiment, the temporal pattern recognition model uses a three-layer causal convolutional network as its basic coding structure, including intra-stage trajectory coding branches and inter-stage transition coding branches. The intra-stage trajectory coding branches extract the growth preservation degree, fluctuation clustering degree, and terminal decay trend of each stage according to unified stage coordinates. The inter-stage transition coding branches extract the duration of the adaptation stage, the steepness of the initiation transition, and the degree of multivariate synchronization according to the stage sequence relationship. Stage boundary confidence is generated based on the continuity of the trajectories on both sides of the adjacent stage boundary, and gating fusion is performed using the stage boundary confidence as a gating signal. The gating fusion uses the stage boundary confidence as the fusion weight of the inter-stage transition coding branch and the complementary value of 1 minus the stage boundary confidence as the fusion weight of the intra-stage trajectory coding branch. The outputs of the two branches are summed element-wise according to their corresponding weights, and stage sequence constraints and non-reverse growth trend constraints are applied at the output.

[0048] The temporal pattern recognition model employs a 3-layer causal convolutional network as its basic encoding structure, achieving causal masking through causal masking. Each convolutional layer contains 32 hidden units, with a kernel size of 3 and dilation coefficients of 1, 2, and 4 layer by layer. The convolution operation at the current time step only utilizes data from the current and historical time steps. The model is divided into intra-stage trajectory encoding branches and inter-stage transition encoding branches. The intra-stage trajectory encoding branch divides the unified biological process coordinates into continuous sub-intervals, calculating the local trend vector, discrete vector, and terminal state vector of the multivariate trajectory within each sub-interval, and inputting them into the causal convolutional network to generate intra-stage trajectory representations. The inter-stage transition encoding branch takes 4 stage nodes and 3 adjacent stage connections as inputs. Stage nodes include the duration, growth preservation, and stable state of the corresponding stage, while stage connections include boundary confidence, multivariate synchronization degree, and transition integrity. Message passing is executed in a fixed order of adaptation, initiation, expansion, and stabilization stages, prohibiting the generation of backward connections from subsequent stages to preceding stages, and outputting inter-stage transition representations. The stage boundary confidence is calculated based on the Pearson correlation coefficient between the trajectories on both sides of the adjacent stage boundary, with a value ranging from 0 to 1. Gated fusion uses the stage boundary confidence as the gating signal. The fusion weight representing the inter-stage transition is equal to the stage boundary confidence, and the fusion weight representing the trajectory within a stage is equal to the difference between 1 and the stage boundary confidence. The fusion feature is obtained by summing the outputs of the two branches element by element according to their corresponding weights.

[0049] Model training uses an objective function with constraints on stage order and growth trend: ; in, This represents the model training objective, and N represents the number of bacterial samples used for training. This represents the known quality characterization value of the m-th sample. This represents the model output value. This represents the sequential state of the m-th sample between adjacent stages. This represents the sequential state of the m-th sample in the next stage. This represents the principal growth trend of the m-th sample in the j-th stage. and These represent the weights of the order constraint and the trend constraint, respectively. The quality characterization values ​​of the training data were labeled by professional inspectors according to the quality inspection procedures for edible fungi strains, including two dimensions: germination ability and sustained vigor, with values ​​ranging from 0 to 1, and normalized with reference to the test results of standard strains in the same batch. The model was trained using the Adam optimization algorithm, with an initial learning rate of 0.001, a learning rate decay coefficient of 0.8 every 20 epochs, a training batch size of 16, and a total of 100 training epochs. Training was terminated early if the validation set loss did not decrease for 10 consecutive epochs. (Weights of the order constraint are listed below.) Values ​​range from 0.1 to 0.5, representing the weight of the trend constraint term. The value ranges from 0.05 to 0.3, and the value is positively adjusted according to the completeness of the labels in the training set stage.

[0050] When the extended phase state given by the model is placed before the initiation phase, the sequence constraint term produces a non-zero value. When the normal growth interval is encoded as a continuous reverse change, the trend constraint term produces a non-zero value. This restricts the output path that does not conform to the continuous growth process of wood-rotting edible fungi during the training process.

[0051] In one embodiment, the consistency determination of the stage state path includes: representing the adaptation stage, startup stage, expansion stage, and stable stage as ordered state nodes, constructing an ordered state diagram that only allows state to remain or transition sequentially along the adaptation stage, startup stage, expansion stage, and stable stage; generating state transition vectors based on the encoding results of adjacent stages, mapping the state transition vectors to the ordered state diagram, identifying conflict segments corresponding to premature decay, repeated startup, and cross-stage reverse transition, and generating path conflict representations based on the stage, duration interval, and stage boundary confidence of the conflict segments.

[0052] In an ordered state diagram, state preservation indicates that the bacterial species remain in the same biological stage at consecutive sampling times. Sequential transitions represent the transition from the adaptation stage to the initiation stage, from the initiation stage to the expansion stage, or from the expansion stage to the stable stage. No conventional connections are established in the diagram to return from the stable stage to the expansion stage or from the expansion stage to the initiation stage. The state transition vector consists of the intra-stage coding differences, inter-stage transition coding results, and boundary confidence at adjacent times, and is mapped to the ordered state diagram according to the principle of maximum consecutive affiliation. If a continuous negative trend occurs before the stable stage and does not subsequently recover to the original growth state, an early decay conflict segment is formed. If, after the germination initiation node, the same high fluctuations as the first initiation occur again, transitioning to a positive cooperative change structure, a repeated initiation conflict segment is formed. If the mapping result requires a return from the subsequent stage to the preceding stage to explain the trajectory, a cross-stage reverse transition conflict segment is formed.

[0053] Path conflict characterization includes conflict type, conflict start and end locations, biological stage, duration interval length, set of involved variables, and corresponding stage boundary confidence. When a conflict segment crosses a stage boundary, the conflict segment is split into an intra-stage part and a boundary part based on the continuity of the trajectories on both sides of the boundary. The intra-stage part is processed by the corresponding stage coding branch, and the boundary part is processed by the inter-stage transition coding branch. Short-term state jumps caused by missing data are not directly identified as path conflicts, but are judged in combination with data integrity markers and the degree of attribution of the preceding and following states. When the two sides of the missing interval still maintain the same state, the intermediate state is regarded as an uncertain state. When a clear sequential transition is formed on both sides of the missing interval, the transition position is restricted to the missing interval and the boundary confidence is reduced.

[0054] In one embodiment, the path conflict characterization, germination ability prediction, and strain vigor evaluation value are input into a consistency calibration layer. The consistency calibration layer encodes conflict segments before the initiation phase as germination calibration components that only affect the germination ability dimension, encodes conflict segments in the expansion and stabilization phases as sustained vigor calibration components that only affect the sustained vigor dimension, and encodes cross-phase reverse transfer as joint constraint components that affect both dimensions simultaneously. The values ​​of the three calibration components range from -1 to 0, and their values ​​are positively correlated with the duration of the conflict segment. The three calibration components are combined according to the node confidence and phase boundary confidence corresponding to each conflict segment, and the combined calibration components are jointly encoded with the output of the temporal pattern recognition model as the strain quality evaluation result.

[0055] The consistency calibration layer generates the calibrated two-dimensional evaluation vector according to the following formula: ; Where r represents the calibrated germination ability and sustained vitality evaluation vector, and y represents the original two-dimensional evaluation vector output by the time-series pattern recognition model. The germination calibration vector is generated from conflict fragments encoded before the initiation phase and has a dimension of 2. This is the vitality calibration vector, generated by encoding conflict fragments from the expansion and stabilization phases. It takes negative values ​​only in the sustained vitality dimension, ranging from -1 to 0. The joint constraint vector is generated by cross-stage reverse transition encoding, with both dimensions taking negative values ​​ranging from -1 to 0; weight coefficients. , and A linear weighting method was used to determine the values ​​of three parameters: node confidence, stage boundary confidence, and data integrity, all within the range of 0 to 1. After calibration, the evaluation vectors were used to classify the levels: a germination ability value greater than or equal to 0.7 was considered high-level, 0.4 to 0.7 was considered medium-level, and less than 0.4 was considered low-level. The same threshold was used for classifying the sustained vitality dimension. When there were conflict segments with long durations and high node confidence before the initiation stage, the effect was mainly on the germination ability dimension; when there was continuous decay during the expansion stage... Primarily affecting the dimension of sustained vitality, when cross-stage reverse transfer involves multiple stages, It acts on two dimensions simultaneously.

[0056] The strain quality evaluation results include germination ability level, sustained vigor level, overall state type, main stage path, and result confidence. The germination ability level is determined based on the relative position of the calibrated germination ability dimension with the reference trajectory of the same batch. The sustained vigor level is determined based on the distance relationship between the calibrated vigor dimension and the high-vigor and low-vigor reference trajectories. The overall state type is divided into at least four categories: rapid start-up and stable expansion, rapid start-up but subsequent decay, slow start-up but stable expansion, and slow start-up with insufficient sustained vigor. The main stage path is recorded according to the actual mapping results in the ordered state diagram. The result confidence is jointly determined by the node confidence, stage boundary confidence, data integrity, and path conflict degree. When any core stage lacks identifiable data, the evaluation result retains the obtained stage conclusions and marks the missing stage as an observation state.

[0057] In one specific embodiment, time-series data of the cultivation process of multiple wood-rotting edible fungi samples from the same batch are continuously acquired. For each sample, five time series are generated: mycelial coverage, edge advancement distance, growth per unit time, mycelial distribution uniformity, and texture density. After trend preservation preprocessing, corresponding first-order change sequences, continuous change sequences, and local stability sequences are generated. In the early stage of cultivation, the coverage and edge advancement distance of a certain sample change little, and the texture density fluctuates locally. Subsequently, four of the five variables show continuous positive changes and maintain cooperative connection in adjacent windows. This time interval is identified as a candidate interval for germination initiation. After the separability test of the preceding and following states, the germination initiation node is determined. In the subsequent trajectory, the growth per unit time changes from large fluctuations to a continuous interval, and the distribution uniformity and the continuity of the growth area stabilize synchronously. The corresponding time interval is identified as a candidate interval for stable expansion, and the stable expansion node is determined.

[0058] Based on the two nodes, the samples are divided into adaptation, initiation, expansion, and stabilization phases, and mapped to a unified biological process coordinate system. In the adaptation phase, duration and fluctuation aggregation are extracted; in the initiation phase, transition steepness and multivariate synchronization are extracted; in the expansion phase, growth retention and local regression are extracted; and in the stabilization phase, terminal decay trend and trajectory dispersion are extracted. Then, stage transition topological features are constructed according to stage nodes and stage connections. Intra-stage trajectory encoding branches handle continuous changes within each stage, while inter-stage transition encoding branches handle the integrity of stage boundary transitions. The two branches perform gated fusion based on stage boundary confidence. The model outputs predicted germination capacity and strain viability evaluation values. An ordered state graph checks the output path and encodes premature decay, repeated initiation, or cross-stage reverse transfer as path conflict representations. The consistency calibration layer then forms the strain quality evaluation result based on this.

[0059] In another preferred embodiment, when the acquisition interval of the time series data of the culture process is not completely consistent between different samples, a continuous time trajectory is first established according to the actual timestamp, and then the resampling position is determined according to the trajectory arc length in the local monotonic interval, so that the interval with slower change uses sparser representation points and the interval with change turning point uses denser representation points. The resampling process must not cross the extreme value preservation mark and stability boundary mark. After the resampling is completed, the first-order change sequence, the continuous change sequence and the local stability sequence are generated. In this way, the time overlap relationship in the cooperative change subplot is determined according to the actual culture time, and the samples with different acquisition intervals can still be compared in the corresponding stages in the biological process coordinate.

[0060] In another preferred embodiment, when the number of complete reference samples in the same culture batch is small, the reference trajectory is composed of samples from historical batches with consistent data semantics and stage node confidence levels that meet the conditions. Before use, intra-batch center trajectory correction is performed. The center trajectory correction only adjusts the baseline position of each variable within the stable interval, without changing the local change direction, the relative order of transition nodes, or the stage duration ratio. When the nodes of the sample to be evaluated are missing, local morphological matching is first performed using samples from the same batch. If samples from the same batch cannot form a concentrated candidate position, historical reference trajectories are introduced, and cross-batch reference sources are encoded into the estimated confidence level, so that the estimated nodes are subject to corresponding restrictions in segmented monotonic mapping and stage boundary fusion.

[0061] In another preferred embodiment, if the strain has not yet formed a stable expansion node when it reaches the predetermined observation endpoint, the complete trajectory after the germination initiation node is treated as an open expansion stage. The boundary of the stable stage is not artificially constructed. The trajectory encoding branch within the stage extracts the growth retention, fluctuation aggregation, and terminal trend of the observed interval. The transition encoding branch between stages only encodes the connection from the adaptation stage to the initiation stage and from the initiation stage to the open expansion stage. The temporal pattern recognition model outputs the germination ability prediction value and the stage vitality evaluation value. The strain quality evaluation result marks the stable stage as an incomplete observation state and retains the interface for updating the stable expansion node and stage state path after subsequent data collection.

[0062] In another preferred embodiment, when some cultivation variables are continuously missing throughout the cultivation process, a collaborative change subgraph is reconstructed for the remaining variables. The number of variables participating in node identification shall not be less than three types of variables that can respectively characterize the spatial expansion state, the temporal growth state, and the organizational stability state, and at least one variable shall be retained for each category. The model training and evaluation use a variable availability mask, so that the trajectory encoding branch within the stage only performs trajectory encoding on the available variables, and the transition encoding branch between stages constructs connections according to the synchronization order between available variables. Missing variables are not replaced with fixed values, and the confidence of the results is limited according to the completeness of the variable categories, so as to maintain the data semantic consistency between the germination initiation node, the stable expansion node, and the stage transition topological features.

[0063] In this embodiment, trend preservation preprocessing, joint temporal characterization, collaborative change subgraph, separability test of preceding and following states, biological process coordinate mapping, stage transition topological features, dual-branch temporal pattern recognition, ordered state path determination, and consistency calibration are performed according to the preceding and following dependencies. Among them, trend preservation preprocessing preserves the true turning points for node identification, node identification provides boundaries for stage alignment, stage alignment provides a unified biological process scale for different samples, stage transition topological features encode the internal state of the stage and the stage transition relationship, stage sequence constraints and non-reverse growth trend constraints limit the state path output by the model, and path consistency calibration maps local high growth, subsequent decay, and stage reverse change to the corresponding evaluation dimensions, thereby forming a complete implementation scheme that can evaluate the vigor and germination ability of wood-rotting edible fungi strains based on the temporal data of the cultivation process.

Claims

1. A method for evaluating the quality of wood-rotting edible fungi strains, characterized in that, include: Obtain multivariate time-series data of the cultivation process of wood-rotting edible fungi strains, perform trend preservation preprocessing on the time-series data of the cultivation process, and generate change sequences; Based on the collaborative change relationship of multiple change sequences within adjacent time windows, the germination initiation node and stable expansion node are determined. Using the germination initiation node and the stable expansion node as anchor points, the absolute culture time is mapped to biological process coordinates and the growth stages are aligned. Extract the stage transition topological features from each aligned growth stage; The stage transition topological features are input into a temporal pattern recognition model constructed based on a 3-layer causal convolutional network, which includes intra-stage trajectory coding branches and inter-stage transition coding branches, and has stage sequence constraints and growth trend constraints. This model obtains germination ability prediction values ​​and strain vigor evaluation values, and generates strain quality evaluation results based on the consistency of stage state paths.

2. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 1, characterized in that, The trend preservation preprocessing includes: constructing an adaptive smoothing window according to the local monotonic intervals of each variable; determining the smoothing intensity based on the directional consistency of adjacent sampling points within the window; suppressing isolated fluctuations where the directional consistency does not meet the continuous change condition; and preserving the original turning point for continuous unidirectional change intervals. Based on the preprocessed time-series data of the cultivation process, first-order change sequences, continuous change sequences, and local stability sequences are formed respectively, and the three types of sequences are combined according to the same time index to form a joint time-series representation for transition identification.

3. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 1, characterized in that, Determining the germination initiation node and the stable expansion node includes: calculating the change direction, duration, and relative increase of each change sequence within a sliding time window, and constructing a directed weighted cooperative change subgraph; the cooperative change subgraph uses each culture variable as a node, and the node attributes include relative increase, duration, local stability, and data integrity; when two culture variables have the same change direction and their duration change intervals overlap, a directed edge is established between the corresponding nodes, and the edge attributes include the overlap time length, the consistency of change direction, and the change initiation order; and a transition candidate interval is formed based on the node retention rate and edge continuation rate of the cooperative change subgraph between consecutive windows. Based on the node preservation relationship and edge connection continuation relationship of the cooperative change subgraph between consecutive windows, a transition candidate interval is formed; Perform a state separability test on the candidate transition intervals. The first candidate transition interval that passes the test is determined as the germination initiation node, and the candidate transition intervals that maintain the growth state for the first time are determined as the stable expansion nodes.

4. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 1, characterized in that, The growth stage alignment includes: dividing each bacterial sample into an adaptation stage, an initiation stage, an expansion stage, and a stable stage based on the germination initiation node and the stable expansion node; For each stage, a segmented monotonic mapping relationship that maintains the time order is established, and the corresponding stages of different bacterial samples are compressed or extended to a unified stage coordinate range. At the stage boundary, a continuous transition weight determined by the node confidence is set, and the trajectory features on both sides of the boundary are fused according to the continuous transition weight to form an aligned temporal trajectory with a unified biological process scale.

5. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 2, characterized in that, The construction of the joint time series representation includes: establishing extremum preservation markers for local extrema in the first-order change sequence, establishing continuous segment markers for continuous segments in the same direction in the continuous change sequence, and establishing stability boundary markers for variance abrupt change locations in the local stability sequence. The three types of markers are encoded into multi-layer trajectory segments according to their temporal adjacency. The multi-layer trajectory segments are then sorted using the synchronous occurrence order of different variables to obtain trajectory feature extraction results that simultaneously include turning points, continuous states, and stable boundaries.

6. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 3, characterized in that, The separability test of the preceding and following states includes: extracting multivariate joint distribution features from the front and back sides of the candidate transition interval, respectively, and constructing state prototypes on both sides of the candidate interval; Calculate the attribution sequence of each time point relative to the two state prototypes, and detect whether the attribution sequence forms a unidirectional continuous transition from the previous state to the next state; Node confidence is generated based on the node retention rate, edge continuation rate, and unidirectional continuous transition length of the cooperative change subgraph, and only the transition candidate intervals that satisfy the cross-window continuity condition are retained.

7. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 4, characterized in that, When a bacterial sample is missing either the germination initiation node or the stable expansion node, a reference trajectory matching the preceding trajectory of the missing node is selected from the bacterial samples in the same batch that have completed node identification. The candidate position of the missing node is inferred based on the local morphological correspondence between the preceding trajectories. The estimated confidence of the candidate positions and the node confidence of the identified nodes are used together to determine the boundary weights of the piecewise monotonic mapping, and the original time proportions are retained for stages where the estimated confidence does not meet the continuous mapping condition.

8. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 4, characterized in that, The temporal pattern recognition model is based on a 3-layer causal convolutional network as its coding structure, including intra-stage trajectory coding branches and inter-stage transition coding branches. The intra-stage trajectory coding branch extracts the growth retention, fluctuation clustering and terminal decay trend of each stage according to the unified stage coordinates; the inter-stage transition coding branch extracts the duration of the adaptation stage, the steepness of the initiation transition and the degree of multivariate synchronization according to the stage sequence. The stage boundary confidence is generated based on the continuity of the trajectories on both sides of the adjacent stage boundary. The stage boundary confidence is used as a gating signal to perform gating fusion. The gating fusion uses the stage boundary confidence as the fusion weight of the inter-stage transition coding branch and the complementary value of 1 minus the stage boundary confidence as the fusion weight of the trajectory coding branch within the stage. The outputs of the two branches are summed element by element according to their corresponding weights, and the stage order constraint and non-reverse growth trend constraint are applied at the output.

9. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 8, characterized in that, The consistency determination of the stage state path includes: representing the adaptation stage, startup stage, expansion stage and stable stage as ordered state nodes, and constructing an ordered state graph that only allows the state to be maintained or to transition sequentially along the adaptation stage, startup stage, expansion stage and stable stage; A state transition vector is generated based on the encoding results of adjacent stages. The state transition vector is then mapped to the ordered state diagram. Conflict segments corresponding to premature decay, repeated startup, and cross-stage reverse transition are identified. Path conflict representations are generated based on the stage, duration interval, and stage boundary confidence of the conflict segments.

10. The method for evaluating the quality of wood-rotting edible fungi strains according to claim 9, characterized in that, The path conflict characterization, the germination ability prediction value, and the strain viability evaluation value are input into the consistency calibration layer; The consistency calibration layer encodes conflict fragments before the initiation phase as germination calibration components that only act on the germination capability dimension, encodes conflict fragments in the expansion and stabilization phases as continuous vitality calibration components that only act on the continuous vitality dimension, and encodes cross-phase reverse transfers as joint constraint components that act on both dimensions simultaneously. The values ​​of the three calibration components all range from -1 to 0, and their values ​​are positively correlated with the duration of conflict segments. The three calibration components are combined according to the node confidence and stage boundary confidence corresponding to each conflict segment, and the combined calibration components are jointly encoded with the output of the time-series pattern recognition model to form the strain quality evaluation result.

Images