Remote monitoring and early warning method for fork mobile robot battery system

Abstract

The application discloses a remote monitoring and early warning method for a fork mobile robot battery system, and particularly relates to the technical field of fork mobile robot battery management; a composite data set is constructed by collecting multi-source operation data, voltage rebound hysteresis characteristics and temperature-current phase offset characteristics are extracted, a non-synchronous characteristic set is formed, and a polarization-diffusion coupled implicit state mapping model is constructed to obtain equivalent representation parameters of unobservable state variables; further, the equivalent representation parameters are mapped to a historical health evolution space to generate a deviated trajectory, and a credible abnormal factor is obtained through reverse consistency verification; an early warning criterion set with a precursor amplification effect is constructed by combining real-time working condition triggering conditions, and a hierarchical early warning result and an adaptive intervention strategy are output by fusing a task scheduling state; the application can realize early identification and dynamic risk control of hidden abnormalities, and improve system operation safety and reliability.

Description

Technical Field

[0001] This invention relates to the field of battery management technology for forklift mobile robots, and more specifically to a remote monitoring and early warning method for battery systems of forklift mobile robots. Background Technology

[0002] Forklift mobile robots (AGV / AMR) operate under conditions of shallow charging and discharging and frequent start-stop in high-frequency handling scenarios in warehousing. The battery system exhibits high dynamics and small fluctuations. Existing remote monitoring methods mostly rely on single voltage and current thresholds for judgment, which makes it difficult to identify hidden anomalies such as "local polarization hysteresis caused by short-term recharging at low temperatures" or "weak reverse degradation of individual cells caused by imbalance mismatch". Especially in unattended nighttime scheduling scenarios, there is a risk of misjudging the health status or delaying the warning, which can lead to extreme risks such as sudden capacity drop or local overheating. Existing methods lack the ability to perform multi-source coupling and time-series correlation analysis. Summary of the Invention

[0003] The purpose of this invention is to provide a remote monitoring and early warning method for battery systems of forklift mobile robots, in order to address the shortcomings in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a remote monitoring and early warning method for a forklift mobile robot battery system, comprising: S100 collects the operating data of the battery system under incomplete charge-discharge cycle conditions and constructs a composite dataset D containing voltage recovery curves, load change response curves, and static decay trajectories. S200 performs cross-condition alignment processing on D, extracts voltage rebound hysteresis features and temperature-current phase shift features at different time scales, and forms a non-synchronous feature set F. S300, based on F, construct an implicit state mapping model H of internal polarization-diffusion coupling of the battery, and obtain the equivalent characterization parameter set Q of the unobservable state quantities; S400, Q is mapped to the historical health evolution space to generate a deviation trajectory, and the deviation trajectory is self-verified by constructing a reverse consistency verification function to obtain the credible anomaly factor C; S500 generates a set of early warning criteria A with a precursor amplification effect based on the nonlinear correlation between C and real-time operating condition triggering conditions. The S600 inputs A into the remote monitoring system and performs coupled corrections based on the robot's task scheduling status, outputting graded early warning results and corresponding adaptive intervention strategy instructions.

[0005] Preferably, the process of constructing the composite dataset D includes: S110 synchronously collects voltage, current and temperature data at a preset sampling frequency during battery system operation, and marks the trigger time of load change to form the original time-series data stream; S120, based on the load change trigger time, segments the original time-series data stream to obtain the corresponding load change response curve, and simultaneously extracts the voltage recovery range after the load is removed. S130 performs curve fitting on the voltage recovery range, extracts characteristic parameters representing the voltage rebound rate, and generates a static decay trajectory by combining the voltage decay data of the static phase. S140, the load mutation response curve, voltage recovery curve and static attenuation trajectory are aligned and fused according to a unified time reference to construct the composite dataset D.

[0006] Preferably, the extraction of voltage rebound hysteresis features and temperature-current phase shift features at different time scales includes: S210, group the various types of data in the composite dataset D according to the trigger time of load change, and perform normalization processing on the time series under different working conditions based on a unified time benchmark to form a comparable standard time series dataset. S220, Based on the standard time series dataset, a multi-time-scale sliding window is constructed, and the voltage recovery curves in each window are differentially calculated to obtain a sequence characterizing the voltage change rate, and the voltage rebound hysteresis characteristic parameters are extracted accordingly. S230, while extracting the voltage rebound hysteresis characteristic parameters, phase matching processing is performed on the temperature data and current data within the corresponding time period to calculate the time offset between temperature change and current change, and the temperature-current phase offset characteristics are obtained.

[0007] Preferably, the extracted voltage rebound hysteresis features are analyzed to generate a voltage rebound hysteresis anomaly index, and the extracted temperature-current phase shift features are analyzed to generate a temperature-current phase shift anomaly index; the voltage rebound hysteresis anomaly index and the temperature-current phase shift anomaly index are combined to form an asynchronous feature set F.

[0008] Preferably, the implicit state mapping model H based on F for internal polarization-diffusion coupling of the battery includes: S310, jointly encodes the voltage rebound hysteresis anomaly index and the temperature-current phase shift anomaly index in the asynchronous feature set F, constructs a multi-dimensional feature vector sequence, and forms the input feature matrix in time order; S320, Based on the input feature matrix, a state mapping function containing polarization response term and diffusion hysteresis term is constructed, and the function parameters are solved by the minimum error iteration method to obtain an implicit state mapping model H describing the feature evolution relationship; S330, Apply the implicit state mapping model H to the current input feature matrix to calculate the corresponding implicit state response sequence, and extract the intermediate state quantities that reflect the polarization intensity and diffusion hindrance. S340 performs normalization and weighted fusion processing on intermediate state variables to generate an equivalent representation parameter set Q for unobservable state variables.

[0009] Preferably, the process of mapping Q to the historical health evolution space to generate a deviation trajectory includes: S410, the polarization equivalent characterization parameters, diffusion equivalent characterization parameters and coupling equivalent characterization parameters in the equivalent characterization parameter set Q are input into the pre-constructed historical health evolution space in chronological order, and the corresponding reference health trajectory points are determined by the minimum distance matching method to form the current state mapping sequence; S420: Based on the difference between the current state mapping sequence and the corresponding reference health trajectory point, calculate the offset of each time point and connect them in chronological order to form the deviation trajectory.

[0010] Preferably, the method for obtaining the credible anomaly factor C includes: S430, based on the deviation trajectory, constructs a reverse consistency verification function, and obtains the trajectory consistency evaluation value by back-projecting the deviation trajectory to the historical health evolution space and calculating the reconstruction error; S440 generates a reliable anomaly factor C based on the combination relationship between trajectory consistency evaluation value and offset.

[0011] Preferably, the generation of the early warning criterion set A with precursor amplification effect includes: S510 collects the real-time operating condition triggering conditions corresponding to the current time point, and performs quantization and encoding processing on the real-time operating condition triggering conditions to form an operating condition feature vector. S520 performs nonlinear coupling calculations between the operating condition feature vector and the reliable anomaly factor C to construct a correlation function containing amplification and suppression factors, and obtains the abnormal response value after operating condition modulation. S530, calculate the precursor amplification factor based on the changing trend of abnormal response values ​​in a continuous time series; S540 combines and maps the abnormal response value with the precursor amplification coefficient to generate a set of early warning criteria A containing multiple judgment thresholds.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs an implicit state mapping model driven by a set of asynchronous features and further introduces a historical health evolution space and a reverse consistency verification function to achieve coupled characterization of the polarization and diffusion behaviors inside the battery. Compared with traditional methods that rely on single voltage and current thresholds or simple statistical features, this invention can identify unobservable early abnormal states without adding additional sensors. By fusing voltage rebound hysteresis anomalies with temperature-current phase shift anomalies across operating conditions and combining the deviation trajectory and its self-verification results to generate a reliable anomaly factor, this invention effectively avoids misjudgment problems caused by instantaneous disturbances and noise. This transforms anomaly identification from "passive threshold triggering" to "active identification based on evolutionary laws," significantly improving the identification accuracy and stability under complex operating conditions such as shallow charging and discharging and high-frequency start-stop.

[0013] 2. This invention introduces a dual-coupling correction mechanism of real-time operating condition triggering conditions and task scheduling status to model the correlation between battery anomaly severity and actual operational risks. It constructs a set of early warning criteria with a precursor amplification effect and outputs tiered early warning results and adaptive intervention strategy instructions. This method can identify potential risks before anomalies reach traditional alarm thresholds and dynamically adjust intervention strategies based on current load status, task priority, and work path, achieving synergistic optimization of the forklift robot's operational safety and work continuity. Compared to existing technologies using fixed thresholds or single-strategy response methods, this invention possesses stronger scenario adaptability and dynamic decision-making capabilities, significantly reducing the risk of sudden failures and improving the overall system reliability and maintenance efficiency. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the remote monitoring and early warning method for a battery system of a forklift mobile robot according to the present invention.

[0016] Figure 2 This is a flowchart of the method for obtaining the reliable anomaly factor of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0018] For examples, please refer to Figure 1 , 2 As shown in this embodiment, the remote monitoring and early warning method for the battery system of a forklift mobile robot includes: S100 collects operating data of the battery system under incomplete charge-discharge cycle conditions and constructs a composite dataset D containing voltage recovery curves, load change response curves, and static decay trajectories.

[0019] During the forklift mobile robot's operations of picking up goods, lifting, turning, braking and retraction, and returning to the warehouse for recharging, a battery management unit installed inside the battery pack continuously collects data on individual battery voltages, total current, and internal battery pack temperature. The battery management unit performs synchronous sampling according to a preset sampling frequency to ensure that voltage, current, and temperature data at the same sampling time have consistent timestamps.

[0020] To ensure that the extracted load change response curve has clear boundaries, the load change trigger time needs to be determined simultaneously with the acquisition of the original time-series data stream. The load change trigger time refers to the moment when the total current undergoes a sudden change between consecutive adjacent sampling points and reaches a preset judgment condition. Specifically, it can be defined as follows: Let the total current at the k-th sampling time be the current value at the k-th sampling time, and the total current at the (k-1)-th sampling time be the current value at the (k-1)-th sampling time. The absolute value of the difference between the two is recorded as the current change amount. When the current change amount is not less than a preset current change threshold, and this change state is maintained continuously for no less than a preset holding time, the first sampling time that meets the condition is marked as the load change trigger time. The preset current change threshold can be defined as 0.15 to 0.30 times the current value corresponding to the rated capacity, and the preset holding time can be defined as 0.3 seconds to 2 seconds.

[0021] The original time-series data stream is constructed as follows: the individual cell voltage, total current, measurement point temperature and timestamp corresponding to each sampling time are arranged in chronological order to form a data sequence with a unified time stamp; then the identified load change trigger time is written into the corresponding timestamp position to form a traceable original time-series data stream.

[0022] After obtaining the load change trigger time, the original time-series data stream is segmented into time windows, centered on each load change trigger time. The time window includes a pre-trigger time period and a post-trigger time period. The pre-trigger time period reflects the steady-state baseline before the load change, and the post-trigger time period reflects the transient response after the load change. Preferably, the pre-trigger time period is set to 2 to 10 seconds, and the post-trigger time period is set to 5 to 30 seconds. Through this segmentation, local data segments corresponding one-to-one with the load change trigger time can be obtained.

[0023] Subsequently, the direction of total current change in the local data segment is determined. A significant increase in total current after the load change trigger indicates that the forklift robot has entered a lifting, acceleration, or load-carrying state; a significant decrease in total current or a shift from discharge to low current after the load change trigger indicates that the external load has been removed or the operation has ended. Based on this determination, voltage and current data before and after the load change trigger are extracted to form a load change response curve. The horizontal axis of the load change response curve represents time relative to the trigger moment, and the vertical axis represents the voltage or current value, thus visually reflecting the transient drop, recovery delay, and fluctuation amplitude of the battery system before and after the load change.

[0024] After extracting the load change response curve, the local data segments after load removal are further processed to obtain the voltage recovery range. The voltage recovery range refers to the time interval during which, after the external load is reduced or removed, the absolute value of the current drops below a preset recovery threshold and remains below it for a preset recovery duration. The preset recovery threshold can be defined as 0.05 to 0.10 times the rated capacity current value, and the preset recovery duration can be defined as 5 to 60 seconds. Using this limitation method, the slow voltage recovery process after load removal can be accurately separated from the overall data, avoiding the miscalculation of data still significantly affected by the load into the voltage recovery range.

[0025] Therefore, the original time-series data stream is segmented at the moment of load change triggering, first obtaining the load change response curve, and then determining the voltage recovery interval based on the current decay in the same time segment, providing basic data for constructing the voltage recovery curve and the static decay trajectory.

[0026] Curve fitting is performed on the obtained voltage recovery interval to generate a voltage recovery curve. The curve fitting is constructed using an exponential recovery function, specifically defined as follows: the voltage value at the recovery moment equals the stable voltage value minus the product of the recovery amplitude and the time-exponential decay term. Here, the stable voltage value represents the voltage level when it finally stabilizes after the load is removed, the recovery amplitude represents the voltage difference between the instant the load is removed and the stable state, and the time decay coefficient represents the voltage recovery rate. These three parameters can be obtained using a least-squares fitting method. The least-squares fitting method is implemented by iteratively solving for the stable voltage value, recovery amplitude, and time decay coefficient, with the objective of minimizing the sum of squared differences between the measured voltage value and the fitted voltage value at each sampling moment within the voltage recovery interval, until the change in the objective function between two adjacent iterations is less than a preset fitting convergence threshold. The preset fitting convergence threshold can be defined as one-thousandth to one-ten-thousandth of the previous objective function value.

[0027] After obtaining the voltage recovery curve, it is necessary to extract characteristic parameters that characterize the voltage rebound rate. These characteristic parameters include at least: the initial recovery slope, the half-recovery time, and the stability deviation. The initial recovery slope refers to the slope of the tangent line at the recovery starting point of the voltage recovery curve, used to characterize the rate of voltage rise after the load is removed; the half-recovery time refers to the time corresponding to the voltage recovering to half of the total recovery amplitude, used to characterize the degree of lag in the recovery process; and the stability deviation is the difference between the fitted final value and the steady-state voltage before the load is removed, used to characterize the degree of residual offset after this recovery.

[0028] A static attenuation trajectory is generated by combining voltage data from the static phase. The static phase refers to a period in which the absolute value of the total current remains below a preset static current threshold for at least a preset static duration. The preset static current threshold can be defined as 0.02 to 0.05 times the current value corresponding to the rated capacity, and the preset static duration can be defined as 10 to 120 minutes. After extracting the static phase voltage data that meets the conditions, the data is arranged in chronological order to obtain a static phase voltage sequence, which is then smoothed using a moving average method to reduce the impact of individual abnormal sampling points. The specific implementation of the moving average method is as follows: a moving window is formed by several consecutive sampling points, and the arithmetic mean of the voltage values ​​within the window is used as the smoothed voltage value at the center moment. The window length can be set to 5 to 30 sampling points. The smoothed static phase voltage sequence constitutes the static attenuation trajectory.

[0029] A voltage recovery curve is generated based on the obtained voltage recovery range, and the voltage rebound rate characteristic parameters are extracted. Simultaneously, a stationary attenuation trajectory is generated using stationary phase data from the same original time-series data stream. The former reflects the recovery capability after dynamic load removal, while the latter reflects the natural attenuation law under low-disturbance conditions.

[0030] The obtained load change response curve, voltage recovery curve, and stationary decay trajectory are transformed using a unified time reference. Since the start times and durations of the three types of data differ, they cannot be directly compared; therefore, a unified time reference is required. The unified time reference is defined as follows: the load change trigger time is set to time zero; the time before and after the trigger in the load change response curve is converted to a relative time relative to time zero; the starting point of the voltage recovery curve is set to the load release time and further mapped to the corresponding relative time zero coordinate; the starting point of the stationary decay trajectory is set to the stationary start time and then converted to the same relative time axis according to the time interval between it and the most recent load change trigger time. This method allows the three types of curves to establish a clear positional relationship under a unified time coordinate.

[0031] Resampling is performed on the three types of data after time base conversion. The purpose of resampling is to eliminate the influence of sampling frequency differences, so that each type of curve has corresponding data points at the same time interval. In specific implementation, a fixed time interval can be selected as the standard sampling interval, preferably 0.1 seconds, 0.5 seconds, or 1 second; for data with an original sampling interval greater than the standard sampling interval, a linear interpolation method can be used to supplement intermediate data points; for data with an original sampling interval less than the standard sampling interval, an interval averaging method can be used for downsampling. The specific implementation of the linear interpolation method is as follows: given the voltage values ​​and timestamps of two consecutive sampling points, the estimated value of the intermediate time is obtained according to the time ratio; the specific implementation of the interval averaging method is as follows: the arithmetic mean of multiple sampling values ​​falling within the same standard time interval is taken.

[0032] After completing the unified time base conversion and resampling process, feature fusion is performed on the load surge response curve, voltage recovery curve, and stationary decay trajectory. This feature fusion is not a simple concatenation, but rather the construction of data recording units according to a unified field format. Each data recording unit includes at least: a relative time value, a corresponding voltage value, a corresponding current value, a corresponding temperature value, a data source category identifier, and the extracted voltage rebound rate feature parameter. The data source category identifier is used to distinguish whether the data belongs to the load surge response curve, voltage recovery curve, or stationary decay trajectory. This structured approach ensures that the composite dataset D simultaneously retains both the original waveform information and the extracted key feature information.

[0033] Finally, the fused data undergoes integrity verification to form a composite dataset D. The integrity verification is performed according to the following rules: if any one of the three data types—the load change response curve, the voltage recovery curve, and the static attenuation trajectory—is missing at a given load change trigger moment, the data is marked as an incomplete sample; if the length of the missing segment in an incomplete sample exceeds a preset missing length threshold, the data is discarded; if the length of the missing segment does not exceed the preset missing length threshold, it is filled in using interpolation results of similar data from adjacent time periods. The preset missing length threshold can be defined as 3 to 10 times the standard sampling interval.

[0034] S200 performs cross-condition alignment processing on D, extracts voltage rebound hysteresis features and temperature-current phase shift features at different time scales, and forms a non-synchronous feature set F.

[0035] First, each data record unit is read from the composite dataset D and grouped according to the load change trigger time. The data record unit corresponding to each load change trigger time is divided into an independent data segment. Each data segment contains complete data of the load change response curve, voltage recovery curve, and static attenuation trajectory under a unified time base.

[0036] Subsequently, normalization is performed on the time series of each data segment. This normalization includes two parts: time normalization and amplitude normalization. Time normalization is implemented by mapping the start time of each data segment to 0, the end time to 1, and mapping intermediate sampling points to the interval [0,1] in a linear proportion, thereby eliminating differences in the duration of different data segments. Amplitude normalization is implemented by using the difference between the maximum and minimum voltage values ​​within the data segment as the normalization benchmark. The voltage value at each sampling point is subtracted from the minimum value and then divided by this difference, so that the normalized voltage values ​​are distributed between 0 and 1. Current data and temperature data are processed in the same way.

[0037] Through the above processing, a standard time-series dataset is obtained for each data segment under a unified time base and a unified amplitude scale. This standard time-series dataset provides consistent input for subsequent multi-time-scale analysis.

[0038] First, a multi-timescale sliding window is constructed on a standard time-series dataset. This multi-timescale sliding window includes three types: short-time windows, medium-time windows, and long-time windows, corresponding to rapid dynamics, moderate changes, and slow recovery processes, respectively. Specifically, the short-time window length is 0.02 to 0.05 of the normalized time axis, the medium-time window length is 0.05 to 0.15, and the long-time window length is 0.15 to 0.30. Each window slides along the time axis with a fixed step size, set to 0.2 to 0.5 of the window length.

[0039] Within each sliding window, differential calculations are performed on the voltage recovery curve. The specific implementation of differential calculations is as follows: for two adjacent sampling points within the window, the difference between the voltage value of the latter sampling point and the voltage value of the former sampling point is calculated and divided by the corresponding time difference to obtain the local voltage change rate; the average value of all differential results within the window is taken to obtain the characteristic value of the voltage change rate of that window.

[0040] Subsequently, the voltage change rate characteristic values ​​obtained at different time scales are combined and analyzed to construct voltage rebound hysteresis characteristic parameters. These parameters include: initial response rate, peak recovery delay time, and rate decay coefficient. The initial response rate is taken as the voltage change rate characteristic value within a short time window; the peak recovery delay time is defined as the normalized time corresponding to the voltage change rate reaching its maximum value; and the rate decay coefficient is obtained by exponentially fitting the voltage change rates of the medium-term and long-term time windows, and its calculation method is to minimize the sum of squared errors between the fitted curve and the actual rate sequence.

[0041] While extracting the voltage rebound hysteresis characteristic parameters, temperature data sequences and current data sequences within the corresponding time period are extracted from the standard time series dataset, ensuring that the two have the same normalized time axis.

[0042] Subsequently, phase matching processing is performed on the temperature and current data sequences. This phase matching processing employs a sliding correlation calculation method, the specific steps of which are as follows: A time offset search interval is set, ranging from -0.2 to 0.2 on the normalized time axis; within this interval, the temperature data sequence is gradually moved with a fixed step size, and the correlation coefficient between the temperature and current data sequences is calculated at each offset. The correlation coefficient is calculated by subtracting the respective mean from the values ​​at corresponding positions in the two sequences, summing the products, and then dividing by the product of their respective standard deviations.

[0043] When the correlation coefficient reaches its maximum value, the corresponding time offset is defined as the time offset of the temperature change relative to the current change. This time offset is used to characterize the degree of lag or advance in the temperature response, thus obtaining the temperature-current phase offset characteristics.

[0044] After obtaining the voltage rebound hysteresis characteristic parameters and temperature-current phase shift characteristics, quantitative analysis was performed on them to generate corresponding anomaly indices.

[0045] First, a voltage rebound hysteresis anomaly index is constructed. This is achieved by selecting voltage rebound hysteresis characteristic parameters from historical normal operating conditions as benchmark samples, and calculating the mean and standard deviation of each characteristic parameter; calculating the difference between the characteristic parameter corresponding to the current data segment and the benchmark mean, and dividing by the standard deviation to obtain the standardized deviation value; and then fusing the standardized deviation values ​​of multiple characteristic parameters using a weighted summation method to obtain the voltage rebound hysteresis anomaly index. The weighting coefficients are pre-set based on the degree of influence of each characteristic parameter on battery performance, with values ​​ranging from 0.2 to 0.5.

[0046] Secondly, a temperature-current phase shift anomaly index is constructed. This is achieved by using the time shift under historical normal operating conditions as a benchmark value, calculating the absolute difference between the current time shift and the benchmark value, and mapping it to the 0-1 interval according to a preset normalization function. The normalization function is implemented using a linear piecewise function. When the difference is less than a first threshold, a low-risk value is output; when the difference is greater than a second threshold, a high-risk value is output. The intermediate interval changes linearly proportionally. The first threshold can be set to 0.02, and the second threshold can be set to 0.10.

[0047] Finally, the voltage rebound hysteresis anomaly index and the temperature-current phase shift anomaly index are combined according to a unified data format to form an asynchronous feature set F. Each element in the asynchronous feature set F contains the voltage rebound hysteresis anomaly index, the temperature-current phase shift anomaly index, and their time labels for the corresponding time period, thus providing a unified input for subsequent implicit state analysis.

[0048] S300, based on F, constructs an implicit state mapping model H of internal polarization-diffusion coupling of the battery, and obtains the equivalent characterization parameter set Q of the unobservable state variables.

[0049] In the preceding steps, a time-stamped set of asynchronous features F has been obtained. Each feature element in the asynchronous feature set F includes at least: the voltage rebound hysteresis anomaly index, the temperature-current phase shift anomaly index, and the time stamp corresponding to the current time period. Since the subsequent construction of the implicit state mapping model H needs to simultaneously reflect the continuous evolution relationship of various anomaly features over time, the asynchronous feature set F first needs to be jointly encoded.

[0050] Specifically, firstly, all feature elements in the asynchronous feature set F are sorted in ascending order according to time labels, so that the abnormal features of adjacent time periods maintain their original temporal sequence. Then, using a fixed-length continuous time period as an encoding unit, the voltage rebound hysteresis anomaly index and temperature-current phase shift anomaly index corresponding to each time point within each encoding unit are combined to form a multi-dimensional feature vector. Preferably, each multi-dimensional feature vector consists of the following components: the voltage rebound hysteresis anomaly index at the current time, the change in the voltage rebound hysteresis anomaly index at the previous time, the temperature-current phase shift anomaly index at the current time, and the change in the temperature-current phase shift anomaly index at the previous time. This construction method allows the input quantity to include not only the degree of anomaly at a certain time but also the trend of the anomaly degree between adjacent time periods. Specifically, the change in the voltage rebound hysteresis anomaly index at the previous time is defined as the voltage rebound hysteresis anomaly index at the current time minus the voltage rebound hysteresis anomaly index at the previous time; the change in the temperature-current phase shift anomaly index at the previous time is defined as the temperature-current phase shift anomaly index at the current time minus the temperature-current phase shift anomaly index at the previous time. If the current time is the first time point, the corresponding change is set to zero to ensure that all multidimensional feature vectors have the same dimensions.

[0051] After obtaining the multidimensional feature vectors at each time point, multiple consecutive multidimensional feature vectors are stacked in chronological order to form an input feature matrix. The rows of the input feature matrix correspond to consecutive time points, and the columns correspond to feature components. Preferably, the input feature matrix is ​​generated using a fixed time window, with the window length set to 5 to 20 consecutive time points; when a new time point enters, the earliest time point is removed, thus forming a sliding update input feature matrix. The purpose of using a fixed time window is to enable the implicit state mapping model H to utilize a continuous historical feature set, rather than just a single moment's feature, thereby more accurately reflecting the cumulative and lagging nature of the polarization and diffusion processes within the battery. To avoid bias in subsequent calculation results due to differences in the range of different anomaly index values, amplitude standardization processing is performed on each feature component before forming the input feature matrix. The specific method of amplitude standardization is as follows: for each column of feature components, the maximum and minimum values ​​in the historical samples are calculated, and mapped to a range between zero and one according to the method of "subtracting the minimum value from the current value and then dividing by the difference between the maximum and minimum values". When the difference between the maximum and minimum values ​​is less than a preset stability threshold, the feature components of that column are directly mapped to a fixed median. The preset stability threshold can be defined as 0.001 to 0.01 to avoid numerical amplification caused by an excessively small denominator.

[0052] Therefore, by performing time sorting, joint encoding, change amount construction, sliding stacking and magnitude normalization on the asynchronous feature set F, an input feature matrix arranged in time order is finally formed, providing a unified input basis for constructing the implicit state mapping model H.

[0053] An implicit state mapping model H is established based on the input feature matrix to reflect the combined effect of polarization and diffusion behavior inside the battery. The specific implementation method is as follows: Because the abnormal performance of the battery under incomplete charge-discharge cycle conditions in a forklift robot is not directly determined by a single voltage or temperature change, but is jointly influenced by charge transfer polarization, concentration diffusion hysteresis, and thermal response delay, this step uses a state mapping function containing polarization response and diffusion hysteresis terms to model the input feature matrix. This state mapping function does not directly provide a physical measurable quantity, but calculates the implicit state response result, which is difficult to obtain directly from sensors, through the mapping relationship between input features and internal states; therefore, it is called the implicit state mapping model H.

[0054] The implicit state mapping model H is constructed as follows: First, let the multidimensional feature vector corresponding to the i-th time point in the input feature matrix be the current feature vector. A polarization response term is constructed based on the current feature vector. This polarization response term characterizes the accumulation of transient anomalies caused by load changes and electrochemical interface charge transfer hysteresis. Its calculation form is defined as follows: multiply the voltage rebound hysteresis anomaly exponent and its change in the current feature vector by their respective weight coefficients, sum them, and then compress them using a monotonically bounded mapping function. The monotonically bounded mapping function is implemented using an exponential compression function, specifically "1 minus the negative exponent term," thus ensuring that the polarization response term always lies between zero and one. This form highlights the growth sensitivity of high anomaly intervals while avoiding state distortion caused by extreme values. A diffusion hysteresis term is constructed based on the current feature vector and the feature vectors of the previous few time points. This diffusion hysteresis term characterizes the delayed anomalies caused by temperature response hysteresis, slow recovery of electrolyte concentration distribution, and the cumulative effect of multiple shallow charge-discharge cycles. The calculation method involves weighted summation of the temperature-current phase shift anomaly indices at the current time point and several previous time points, assigning higher weights to time points closer to the current time point and lower weights to time points further away. The weights are constructed using a decaying sequence, specifically defined as a sequence where the first term is 1, and adjacent terms decrease by a fixed ratio of 0.6 to 0.9. This method preserves the residual effects of historical anomalies.

[0055] After obtaining the polarization response term and the diffusion hysteresis term, they are combined to form a state mapping function. Preferably, the state mapping function is defined as: In the formula, This represents the implicit state response value at time point t. This represents the polarization response term at time point t. This represents the diffusion lag term at time point t. This represents the implicit state response value at time point t-1. This represents the weighting coefficient of the polarization response term. This represents the weighting coefficient of the diffusion lag term. Let α, β, and γ represent the state continuation coefficients. α, β, and γ are all real numbers greater than 0, with γ ranging from 0.3 to 0.8. To ensure the comparability of the contributions of each component to the implicit state response value at the current moment, it is preferable to limit α + β + γ = 1. Under this constraint, α characterizes the immediate impact of the voltage rebound hysteresis anomaly on the current implicit state response value, β characterizes the delayed impact of the temperature-current phase shift anomaly on the current implicit state response value after historical accumulation, and γ characterizes the inertial strength of the continuation of the implicit state response value from the previous moment to the current moment.

[0056] To determine the weight coefficients, state continuity coefficients, and compression function parameters in the state mapping function, it is necessary to solve for these parameters. This implementation method uses the minimum error iteration method. The specific implementation steps of the minimum error iteration method are as follows: Historical known operating condition samples are selected as training samples. These samples include an input feature matrix for the corresponding time period and risk level labels confirmed through long-term operational results. The risk level labels can be obtained by reverse annotation based on subsequent events such as abnormal capacity reduction, localized overheating, or abnormal shutdowns.

[0057] Initial values ​​are assigned to each parameter in the state mapping function. Preferably, the initial values ​​of the polarization response term weight coefficient and the diffusion hysteresis term weight coefficient are set to 0.5, the initial value of the state continuity coefficient is set to 0.5, and the initial value of the compression function parameter is set to 1.

[0058] The training samples are input into a state mapping function to calculate the implicit state response value at each time point, and a predicted risk value is generated accordingly. The predicted risk value can be obtained by averaging or taking the maximum value of the implicit state response values ​​at multiple consecutive time points.

[0059] Calculate the error between the predicted risk value and the risk level label. The error function is in the form of a sum of squared errors, which is the sum of the squares of the predicted risk value minus the label value for all samples.

[0060] Adjust each parameter sequentially in the direction of error reduction until the error change between two consecutive iterations is less than a preset convergence threshold. The preset convergence threshold can be defined as 0.0001 to 0.001. If the convergence condition is not met even after reaching the maximum number of iterations, the set of parameters with the smallest error is selected as the final solution. The maximum number of iterations can be defined as 100 to 1000.

[0061] After the above solution, an implicit state mapping model H that can describe the asynchronous characteristic evolution relationship is obtained. This model does not rely on the direct measurement of the actual polarization resistance or actual diffusion coefficient inside the battery, but indirectly characterizes the internal evolution through anomalous features and state recursion relationships, thus having strong adaptability and protectability.

[0062] After the parameters are solved, the input feature matrix corresponding to the current time period to be analyzed is input into the implicit state mapping model H. The model calculates the implicit state response value for each time point in chronological order, and arranges the implicit state response values ​​of all time points in order to form an implicit state response sequence.

[0063] The implicit state response sequence is used to reflect the degree of anomaly accumulation that is difficult to directly observe inside the battery under continuous operating conditions. To further improve interpretability, intermediate state quantities need to be extracted based on the implicit state response sequence. These intermediate state quantities include at least: polarization intensity indicators and diffusion hindrance indicators.

[0064] The polarization intensity indicator is extracted as follows: during the calculation of the implicit state mapping model H, the values ​​of the polarization response terms at each time point are retained, and then a weighted average is performed on all polarization response term values ​​within the current time window to obtain the polarization intensity indicator. During the weighted average, time points closer to the current time have higher weights, and time points farther from the current time have lower weights, to ensure that the current anomaly has a dominant role in the result.

[0065] The diffusion lag indicator is extracted as follows: During the calculation of the implicit state mapping model H, the values ​​of diffusion lag terms at each time point are retained. Then, all diffusion lag terms within the current time window are summed and divided by the time window length to obtain the diffusion lag indicator. This summation and averaging process reflects the characteristic of diffusion anomalies having a continuous cumulative effect.

[0066] Furthermore, to reflect the rate of state change, the state growth slope can be further calculated from the implicit state response sequence. The state growth slope is defined as the difference between the implicit state response value at the current time point and the implicit state response value at the previous time point, divided by the interval between adjacent time points. If the state growth slope remains positive and exceeds a preset growth threshold, it indicates that internal anomalies are accumulating rapidly. The preset growth threshold can be defined as 0.01 to 0.05. This state growth slope can be used as supplementary intermediate state quantities in subsequent fusion.

[0067] Because the obtained polarization intensity indicators, diffusion hindrance indicators, and state growth slopes differ in numerical range and sensitivity to change, normalization is required before fusion. The normalization process involves calculating the lower and upper limits of each intermediate state value in the historical training samples, and mapping each intermediate state value to a range of 0 to 1 by subtracting the lower limit from the current value and then dividing by the difference between the upper and lower limits. When an intermediate state value exceeds a historical upper limit, it is truncated to 1; when it falls below a historical lower limit, it is truncated to 0. This truncated normalization prevents extreme outliers from compromising the overall fusion stability.

[0068] After normalization, the intermediate state variables are weighted and fused. Preferably, the equivalent characterization parameter set Q includes at least the following parameters: polarization equivalent characterization parameter, diffusion equivalent characterization parameter, and coupling equivalent characterization parameter. The polarization equivalent characterization parameter is directly obtained from the normalized polarization intensity indicator and is used to characterize the degree of anomaly caused by charge transfer lag within the current time window of the battery. The diffusion equivalent characterization parameter is directly obtained from the normalized diffusion retardation indicator and is used to characterize the degree of anomaly caused by slow recovery of internal mass migration. The coupling equivalent characterization parameter is obtained by multiplying the normalized polarization intensity indicator and the normalized diffusion retardation indicator, and then combining this with a weighted correction value for the state growth slope. The weighted correction value can be defined as the state growth slope multiplied by a correction weight, with the correction weight ranging from 0.1 to 0.4. This product-based correction method highlights high-risk states that only occur when polarization and diffusion anomalies coexist and the anomalies are still intensifying.

[0069] Finally, the polarization equivalent characterization parameters, diffusion equivalent characterization parameters, and coupling equivalent characterization parameters are written into the result data record according to a unified field format, forming the equivalent characterization parameter set Q for unobservable state quantities. Since this parameter set does not directly measure the actual polarization resistance, actual diffusion coefficient, or actual concentration distribution inside the battery, but rather represents the equivalent expression after a stable mapping of unobservable states, it can provide continuous, calculable, and physically relevant input quantities for subsequent historical health evolution spatial mapping and deviation trajectory analysis without adding additional sensors.

[0070] S400 maps Q to the historical health evolution space to generate a deviation trajectory, and performs self-verification on the deviation trajectory by constructing a reverse consistency verification function to obtain the credible anomaly factor C.

[0071] In this embodiment, the equivalent characterization parameter set Q includes at least polarization equivalent characterization parameters, diffusion equivalent characterization parameters, and coupling equivalent characterization parameters, which are used to characterize polarization-related anomalies, diffusion-related anomalies, and coupling anomalies caused by the combined effect of the two within the battery at the current time point, respectively. Since the parameter values ​​at a single time point cannot fully reflect the continuous evolution of the health state, before performing mapping, the equivalent characterization parameter set Q corresponding to multiple consecutive time points is first read in chronological order to form a sequence of parameters to be mapped. Preferably, the sequence of parameters to be mapped consists of 5 to 30 consecutive time points, thereby preserving both short-term change characteristics and certain evolutionary trend information.

[0072] The historical health evolution space is constructed from historical operating samples. Its construction method is as follows: First, historical data of batteries in different health stages of a forklift mobile robot during long-term operation are collected; then, the polarization equivalent characterization parameters, diffusion equivalent characterization parameters, and coupling equivalent characterization parameters corresponding to each historical time point are calculated according to the aforementioned steps; subsequently, the three types of parameters corresponding to each historical time point are treated as a three-dimensional state point and connected in chronological order to form one or more historical health trajectories. The historical health trajectories cover the normal health stage, the slow decay stage, and the abnormal degradation stage. The state points in all historical health trajectories are set in a unified coordinate space to form the historical health evolution space.

[0073] To improve the comparability of mapping results, a unified scaling process needs to be applied to the three types of parameters when constructing the historical health evolution space. Specifically, the lower and upper limits of the polarization equivalent characterization parameters, diffusion equivalent characterization parameters, and coupling equivalent characterization parameters in historical samples are statistically calculated, and linear normalization is used to map these three types of parameters to the range of zero to one. The parameter values ​​in the current sequence to be mapped are also normalized using the same lower and upper limits, thus ensuring that the current state and historical states are on the same coordinate scale.

[0074] After completing the above preparations, each time point parameter in the parameter sequence to be mapped is sequentially input into the historical health evolution space, and the corresponding reference health trajectory point is determined using the minimum distance matching method. The specific implementation steps of the minimum distance matching method are as follows: Take a parameter at a certain time point in the parameter sequence to be mapped, and denote it as the current parameter point. This current parameter point consists of the normalized polarization equivalent characterization parameter, diffusion equivalent characterization parameter, and coupling equivalent characterization parameter.

[0075] Calculate the distance between the current parameter point and all historical state points in the historical health evolution space. The distance is defined using weighted Euclidean distance, specifically expressed as: the squares of the differences between the current parameter point and a certain historical state point in each of the three parameter dimensions are multiplied by their respective weights, summed, and then the square root is taken. The weights corresponding to the three dimensions are pre-set according to the parameter sensitivity. Preferably, the weight of the coupling equivalent characterization parameter dimension is greater than that of the polarization equivalent characterization parameter dimension and the diffusion equivalent characterization parameter dimension.

[0076] The historical state point with the smallest distance value is selected from all historical state points and used as the reference health trajectory point corresponding to the current parameter point. If multiple historical state points have distance values ​​less than a preset distance resolution threshold, the historical state point that is chronologically continuous with the reference health trajectory point at the previous time point is selected first to maintain the continuity of the mapped trajectory. The preset distance resolution threshold can be defined as 0.01 to 0.05.

[0077] Repeat the above process for all time points in the sequence of parameters to be mapped, resulting in multiple reference health trajectory points arranged in chronological order. Map these reference health trajectory points one-to-one with the current parameter points to form the current state mapping sequence.

[0078] In the steps described above, a corresponding reference health trajectory point has been found for each time point in the parameter sequence to be mapped. Therefore, in this step, the offset between the current parameter point and the corresponding reference health trajectory point at each time point is first calculated. The offset reflects not only the distance between the current position and the reference position, but also the direction of deviation in each parameter dimension.

[0079] Specifically, the offset at a given time point consists of an offset vector and an offset magnitude. The offset vector is defined as the combination of differences between the current parameter point and the reference healthy trajectory point in the dimensions of polarization equivalent characterization parameters, diffusion equivalent characterization parameters, and coupling equivalent characterization parameters; the offset magnitude is defined as the magnitude of the offset vector. This definition allows for the simultaneous acquisition of both the degree and structure of the current state's offset relative to the reference healthy trajectory point.

[0080] To enhance the stability of subsequent analysis, the offset can be smoothed. The smoothing is achieved using a moving average method, specifically: taking the current time point as the center, selecting the offset amplitudes of several adjacent time points, calculating their arithmetic mean, and using this average as the smoothed offset amplitude for the current time point. Preferably, the sliding window length is set to 3 to 7 time points. If the current time point is located at the sequence boundary, the average value of the actual available time points is used. This processing method can reduce isolated spikes caused by single instantaneous perturbations.

[0081] After obtaining the offsets at each time point, the endpoints of each offset vector are connected sequentially in chronological order to form the deviation trajectory. The deviation trajectory can be understood as the offset process of the actual movement path of the current battery state in the historical health evolution space relative to the most recent reference healthy path. If the deviation trajectory continuously deviates along a certain parameter dimension at multiple consecutive time points, it indicates that the current battery state is evolving in a direction dominated by that type of anomaly; if the deviation trajectory exhibits oscillations with small offset amplitudes, it indicates that the current state is still within a recoverable fluctuation range.

[0082] To facilitate quantitative analysis of the deviation trajectory, trajectory feature parameters can be further extracted. These trajectory feature parameters include at least: cumulative trajectory length, average trajectory offset amplitude, and trajectory direction change rate. The cumulative trajectory length is defined as the sum of the distances between the endpoints of the offset vectors at adjacent time points, used to characterize the overall activity level of the deviation process; the average trajectory offset amplitude is defined as the average of the smoothed offset amplitudes at all time points, used to characterize the overall degree of deviation; and the trajectory direction change rate is defined as the average of the changes in the angle between two adjacent offset vectors, used to characterize whether the deviation direction is stable. A smaller trajectory direction change rate indicates a more concentrated deviation direction; a larger rate of change indicates a less stable deviation direction.

[0083] Therefore, by calculating the offset between the current state mapping sequence and the corresponding reference health trajectory point, and connecting them in chronological order to form a deviation trajectory, the offset process of the battery's current health state relative to the historical health evolution trend can be continuously described, providing input for subsequent self-verification calculations.

[0084] Since the deviation trajectory is composed of the difference between the current parameter point and the reference healthy trajectory point, if there is instantaneous noise, local sampling jitter, or non-continuous anomalies caused by a single operation in the input data, it may cause a short-term shift in the deviation trajectory. However, such shifts may not accurately reflect the direction of healthy evolution. Therefore, this step constructs a reverse consistency verification function to perform back projection and error evaluation on the deviation trajectory.

[0085] The core idea of ​​the reverse consistency verification function is: if the deviation trajectory does indeed originate from the real deviation of the current state relative to the historical healthy evolution space, then after the deviation trajectory is back-projected to the historical healthy evolution space, it should be able to reconstruct a parameter change path that is similar to the original current state mapping sequence; conversely, if the deviation trajectory mainly originates from accidental disturbances, then there will be a significant difference between the reconstructed path after its back projection and the original current state mapping sequence.

[0086] The steps for constructing the reverse consistency verification function are as follows: Using the obtained reference health trajectory points as reference points and the offset vectors at the corresponding time points as correction values, a reverse superposition operation is performed on each reference health trajectory point to generate reconstructed state points. Specifically, the value of the reconstructed state point in each parameter dimension is equal to the parameter value of the corresponding reference health trajectory point in that dimension plus the component of the offset vector at the same time point in that dimension.

[0087] The reconstructed state points corresponding to all time points are connected in chronological order to form a reconstructed trajectory. The reconstructed trajectory is used to simulate "the current state path recovered by adding the historical healthy path to the deviation trajectory".

[0088] The reconstructed trajectory is compared point-by-point with the original parameter sequence to be mapped, and the reconstruction error is calculated. The reconstruction error is defined using a weighted sum of squared errors, which is calculated by squaring the difference between the reconstructed state point parameter value and the original current parameter point parameter value at each time point and for each parameter dimension, multiplying it by the corresponding dimension weight, and summing the results. This summation is applied to all time points. The weights are consistent with the weighted Euclidean distance weights to ensure a uniform evaluation scale.

[0089] To avoid incomparability of error values ​​between trajectories of different lengths, the reconstruction error is normalized. Specifically, the reconstruction error is divided by the number of time points contained in the trajectory to obtain the average reconstruction error per unit time point.

[0090] The trajectory consistency evaluation value is generated based on the average reconstruction error per unit time point. Preferably, the trajectory consistency evaluation value is calculated using a reverse mapping function, specifically defined as "one divided by one plus the average reconstruction error per unit time point". Using this form, the trajectory consistency evaluation value always lies between zero and one, and the smaller the reconstruction error, the closer the trajectory consistency evaluation value is to one; the larger the reconstruction error, the closer the trajectory consistency evaluation value is to zero.

[0091] In a preferred embodiment, a consistency judgment threshold can be set to preliminarily determine whether the trajectory has passed self-verification. The consistency judgment threshold can be defined as 0.6 to 0.85. When the trajectory consistency evaluation value is not less than the consistency judgment threshold, the deviation trajectory is determined to have high consistency; when the trajectory consistency evaluation value is less than the consistency judgment threshold, the deviation trajectory is determined to have high uncertainty. This judgment result does not directly output a warning conclusion, but rather serves as a verification input in the subsequent generation of the reliable anomaly factor C.

[0092] The deviation index preferably uses a weighted combination of the average trajectory deviation amplitude and the cumulative trajectory length as the basic anomaly quantity. The calculation method is as follows: multiply the average trajectory deviation amplitude by a first anomaly weight, multiply the cumulative trajectory length by a second anomaly weight, and then sum the two to obtain the basic anomaly quantity. Both the first and second anomaly weights are real numbers greater than zero, and their sum is preferably 1. If the importance of continuous deviation is emphasized, the second anomaly weight is greater than the first anomaly weight; if the importance of immediate deviation intensity is emphasized, the first anomaly weight is greater than the second anomaly weight.

[0093] After obtaining the basic anomaly quantity, it is coupled with the trajectory consistency evaluation value to generate a reliable anomaly factor C. Preferably, the reliable anomaly factor C is calculated by multiplying the basic anomaly quantity by the trajectory consistency evaluation value, and then adjusting it by incorporating a correction term for the trajectory direction change rate. The reason for introducing the correction term for the trajectory direction change rate is that when the deviation direction is highly discrete, the deviation trajectory is more likely to originate from random fluctuations; when the deviation direction is continuously concentrated, the deviation trajectory is more likely to originate from the evolution of real anomalies. Therefore, the trajectory direction change rate can be used as a reverse correction quantity. Specifically, the trajectory direction change rate is first normalized, and then "1 minus the normalized trajectory direction change rate" is used as the direction stability coefficient. Finally, the reliable anomaly factor C can be jointly determined by the basic anomaly quantity, the trajectory consistency evaluation value, and the direction stability coefficient.

[0094] To facilitate subsequent calls, the trusted anomaly factor C can be bounded. Specifically, when the trusted anomaly factor C is less than zero, it is truncated to zero; when the trusted anomaly factor C is greater than one, it is truncated to one. This process ensures that the trusted anomaly factor C stably falls between zero and one. The closer the value is to one, the greater the deviation of the current anomaly, and the higher the consistency and directional stability of the deviation trajectory. The closer the value is to zero, the lower the deviation of the current anomaly, or the lack of sufficient self-checking support.

[0095] In a preferred embodiment, the credibility anomaly factor C can be further classified. Specifically, when the credibility anomaly factor C is less than a first credibility threshold, it is determined to be a low-credibility anomaly; when the credibility anomaly factor C is between the first and second credibility thresholds, it is determined to be a medium-credibility anomaly; and when the credibility anomaly factor C is greater than the second credibility threshold, it is determined to be a high-credibility anomaly. The first credibility threshold can be defined as 0.3 to 0.5, and the second credibility threshold can be defined as 0.6 to 0.8. This classification result can serve as the input basis for subsequently generating the warning criterion set A.

[0096] Therefore, by fusing the magnitude information of the deviation trajectory, the trajectory consistency evaluation value, and the degree of directional stability, a reliable anomaly factor C is generated. This enables the anomaly characterization results to reflect not only the magnitude of the deviation but also the reliability and persistence of the deviation, thereby providing a more discriminative input for the construction of subsequent early warning criteria.

[0097] S500 generates a set of early warning criteria A with a precursor amplification effect based on the nonlinear correlation between C and real-time operating condition triggering conditions.

[0098] In this embodiment, the real-time operating condition triggering condition refers to operating condition information corresponding to the current time point and capable of promoting or mitigating abnormal battery evolution. The real-time operating condition triggering condition includes at least: current load intensity, lifting action duration, frequency of continuous start-stop cycles, number of recharging cycles per unit time, ambient temperature range, interval between resuming operation after charging, and the urgency of the vehicle's current task. The above operating conditions are selected because in high-frequency handling scenarios, battery abnormalities in forklift mobile robots are not solely determined by internal conditions but are also affected by external load impacts, repeated short-interval recharging, low-temperature lifting operations, and continuous heavy-load operations.

[0099] To ensure that different types of operating conditions can participate in calculations within the same computational framework, it is first necessary to quantize and encode the triggering conditions for each real-time operating condition. The specific steps of the quantization and encoding process are as follows: The first step is to collect the raw values ​​of various real-time operating condition triggering conditions corresponding to the current time point. Among them, the current load intensity can be characterized by the average and peak values ​​of traction current and lifting current within the current time window; the duration of the lifting action can be characterized by the duration of the lifting execution signal; the frequency of continuous start and stop can be characterized by the number of drive direction switching times and the number of times the speed rises from zero to the set value per unit time; the number of charging times per unit time can be characterized by the number of connected charging actions that occurred in the most recent preset time period; the ambient temperature range can be characterized by the temperature of the measuring points in the battery compartment; the interval between resuming operation after charging is completed can be characterized by the difference between the time when charging stops and the time when the load state is re-entered; and the urgency of the vehicle's current task can be characterized by the scheduling task priority and the number of tasks to be executed per unit time.

[0100] The second step is to divide the trigger conditions for each real-time operating condition into intervals. The purpose of interval division is to transform continuous variables into risk level inputs with clear physical meaning. Taking the current load intensity as an example, it can be divided into low load intervals, medium load intervals, and high load intervals according to the proportion of rated operating current; taking the frequency of continuous start-stop as an example, it can be divided into low frequency intervals, medium frequency intervals, and high frequency intervals according to the number of start-stops within a unit of ten minutes; taking the interval length between restarting operation after charging as an example, it can be divided into short interval intervals, medium interval intervals, and long interval intervals. The interval boundaries are set by statistically analyzing the distribution positions of corresponding operating condition values ​​in historical normal samples and historical abnormal samples, and using the quantiles that can significantly distinguish between the two types of samples as boundaries. Preferably, the 30th and 70th percentiles of the historical distribution can be used as initial boundaries, and then fine-tuned based on the actual sample differentiation effect.

[0101] The third step is to quantify and assign values ​​to the various real-time operating condition triggering conditions after the interval division is completed. Preferably, for operating condition intervals that promote the spread of anomalies, an increasing positive quantization value is assigned; for operating condition intervals that are conducive to anomaly recovery or unfavorable to anomaly spread, a lower quantization value or a suppressive quantization value is assigned. Taking the ambient temperature interval as an example, a lower quantization value is assigned within the suitable temperature range, and a higher quantization value is assigned within the low temperature heavy load interval; taking the interval between resuming operation after charging as an example, a higher quantization value is assigned to short interval intervals, and a lower quantization value is assigned to long interval intervals.

[0102] The fourth step involves arranging the quantized values ​​corresponding to each real-time operating condition trigger condition in a preset order to form an operating condition feature vector. Each component in the operating condition feature vector corresponds to a specific operating condition, and their arrangement order remains consistent between the training and application phases. To eliminate the impact of inconsistent numerical ranges among the components, amplitude normalization processing is required after forming the operating condition feature vector. Amplitude normalization employs a linear mapping method to uniformly transform each component to a value between zero and one. If a component is an inhibitory quantized value, it is first shifted before linear mapping to ensure that all components are within a uniform numerical range.

[0103] In this implementation, the credible anomaly factor C reflects the reliability and persistence of the current anomaly deviation, but it does not directly indicate whether the anomaly will rapidly expand under specific operating conditions. Therefore, the credible anomaly factor C and the operating condition feature vector need to be input into the correlation function together to obtain an anomaly response value modulated by the operating conditions that better reflects the characteristics of on-site risk propagation.

[0104] The correlation function consists of two parts: an amplification factor and an inhibition factor. The amplification factor characterizes the promoting effect of certain operating conditions on the spread of anomalies; the inhibition factor characterizes the mitigating effect of certain operating conditions on the spread of anomalies. The specific implementation is as follows: First, components significantly correlated with abnormal amplification are selected from the operating condition feature vector to construct the amplification factor. Preferably, the current load intensity, frequency of continuous start-stop operations, number of recharging operations per unit time, and short-interval re-entry characteristics are included in the amplification factor calculation. The amplification factor can be defined as the sum of the products of each amplified operating condition component and its corresponding weight, obtained by mapping with an exponential gain function. Its calculation expression can be written as: ;in, This represents the magnification factor at time point t. This represents the i-th amplified type of operating condition component in the operating condition feature vector at time t. denoted by , n represents the weighting coefficient corresponding to the i-th amplification category component; n represents the number of amplification category components; and μ represents the upper limit coefficient of amplification gain. The upper limit coefficient of amplification gain is preferably between 0.5 and 2.0. Using this form, as the amplification category component increases, the amplification factor monotonically increases and becomes greater than 1, thus reflecting the early enhancement effect of the operating condition on the anomaly.

[0105] Secondly, components significantly correlated with anomaly mitigation are selected from the operating condition feature vector to construct a suppression factor. Preferably, a relatively long interval between reactivation after charging and a suitable ambient temperature range are included in the suppression factor calculation. The suppression factor can be defined as the sum of the products of the suppression-type operating condition components and their corresponding weights, obtained through an attenuation mapping function. Its calculation expression can be written as: ;in, This represents the inhibition factor at time point t; This represents the j-th suppressed operating condition component in the operating condition feature vector at time t. ν represents the weighting coefficient corresponding to the j-th suppression-type operating condition component; m represents the number of suppression-type operating condition components; and ν represents the suppression adjustment coefficient. The suppression adjustment coefficient is preferably between 0.5 and 3.0. Using this form, the stronger the suppression-type operating condition component, the closer the suppression factor is to a smaller value, thereby weakening the output results of abnormal expansion.

[0106] After obtaining the amplification factor and suppression factor, they are coupled with the reliable anomaly factor C to obtain the abnormal response value after operating condition modulation. Its calculation expression can be defined as: ;in, This represents the abnormal response value after operating condition modulation at time point t; Let C represent the credible anomaly factor at time point t.

[0107] In this embodiment, the abnormal response value after operating condition modulation at a single time point can only indicate the risk level at that moment, but cannot indicate whether the risk is an instantaneous spike or a precursor to continuous expansion. Therefore, it is necessary to analyze the growth trend of the abnormal response value after operating condition modulation on a continuous time series, and then extract the precursor amplification factor.

[0108] Specifically, firstly, bounded, condition-modulated abnormal response values ​​corresponding to multiple consecutive time points are read in chronological order to form an abnormal response sequence. Preferably, the abnormal response sequence consists of 5 to 20 consecutive time points. Then, trend analysis processing is performed on the abnormal response sequence.

[0109] The trend analysis process includes three parts: slope calculation, continuous growth counting, and local acceleration calculation.

[0110] The first part is the slope calculation. The specific method for slope calculation is as follows: For two adjacent time points in the abnormal response sequence, calculate the bounded, modulated abnormal response value of the later time point minus the bounded, modulated abnormal response value of the previous time point, and divide by the time interval between the two time points to obtain the local growth slope. Then, average all local growth slopes within the current analysis window to obtain the average growth slope. The larger the average growth slope, the more obvious the anomalous amplification trend.

[0111] The second part is the continuous growth count. The specific method for continuous growth counting is as follows: starting from the current time point and proceeding backwards, determine whether the bounded, modulated abnormal response value at each adjacent time point continuously increases; if the value at the later time point is greater than the value at the previous time point, increment the count; if the condition for continuous increase is not met, stop counting. This ultimately yields the continuous growth length corresponding to the current time point. A larger continuous growth length indicates that the abnormality is not an isolated fluctuation, but rather a continuously expanding phenomenon.

[0112] The third part is the calculation of local acceleration. The specific method for calculating local acceleration is as follows: after obtaining the local growth slope at adjacent time points, calculate the difference between two adjacent local growth slopes and divide it by the corresponding time interval to obtain the local growth acceleration; average all local growth accelerations within the current analysis window to obtain the average growth acceleration. An average growth acceleration greater than zero indicates that the abnormal growth rate itself is still accelerating.

[0113] After completing the calculations for the above three parts, the average growth slope, continuous growth length, and average growth acceleration are fused to generate the precursor amplification factor. Preferably, the expression for calculating the precursor amplification factor is: ;in, This represents the precursor amplification factor at time point t; The normalized value representing the average growth slope; Normalized value representing the length of continuous growth; This represents the normalized value of the average growth acceleration. These represent the corresponding weighting coefficients. Preferably, =1, and all three are real numbers greater than zero.

[0114] To facilitate comparison, the average growth slope, continuous growth length, and average growth acceleration need to be normalized first. The normalization process involves calculating the lower and upper limits of these three values ​​from historical samples and then converting them to values ​​between 0 and 1 using a linear mapping method. Values ​​exceeding the upper limit are truncated to 1; values ​​below the lower limit are truncated to 0.

[0115] In a preferred embodiment, a precursor amplification judgment threshold can also be set to determine whether the current anomaly has obvious precursor expansion characteristics. The precursor amplification judgment threshold can be defined as 0.4 to 0.7. When the precursor amplification coefficient at time point t is not less than the precursor amplification judgment threshold, the current anomaly is determined to have a precursor amplification effect; when it is less than the threshold, the current anomaly is determined not to have formed an obvious and continuous amplification trend.

[0116] In this embodiment, the abnormal response value after operating condition modulation mainly reflects the intensity of the abnormality under the current operating condition, while the precursor amplification coefficient mainly reflects whether the abnormality is continuously expanding. Using only one of these factors is insufficient to fully characterize the true risk; therefore, it is necessary to jointly map the two to construct a set of early warning criteria A that can cover different risk levels.

[0117] Specifically, the bounded, modulated abnormal response value at the current time point is first combined with the corresponding precursor amplification coefficient to form a two-dimensional judgment point. The bounded, modulated abnormal response value is used as the first judgment dimension, and the precursor amplification coefficient is used as the second judgment dimension. Then, multiple early warning zones are pre-divided in this two-dimensional judgment space, with different early warning zones corresponding to different risk levels.

[0118] The early warning area is constructed as follows: Historical operational samples are selected as criteria for sample construction. These historical operational samples include normal operation samples, slightly abnormal samples, and severely abnormal samples, and each sample corresponds to a real consequence label, such as whether a sudden drop in capacity occurred, whether a sustained local temperature rise occurred, or whether an operation was interrupted. The bounded, condition-modulated abnormal response values ​​and precursor amplification coefficients of each historical sample at corresponding time points are plotted in a two-dimensional judgment space to form a sample distribution point set. Based on the sample distribution point set, multi-level judgment threshold boundaries are determined. Preferably, a first warning threshold boundary, a second warning threshold boundary, and a third warning threshold boundary can be set. The first warning threshold boundary is used to distinguish between normal risk and low-level warning risk; the second warning threshold boundary is used to distinguish between low-level warning risk and medium-level warning risk; and the third warning threshold boundary is used to distinguish between medium-level warning risk and high-level warning risk. The threshold boundaries are determined by searching for the classification accuracy corresponding to different boundary combinations while ensuring that high-risk samples are located in high-level regions as much as possible and normal samples are located in low-level regions as much as possible. The set of boundaries with the highest overall accuracy is selected as the final threshold boundaries.

[0119] In a preferred embodiment, the threshold boundaries at each level can be defined using a piecewise linear function. Specifically, it can be written as: ;in, , and These represent the first, second, and third warning threshold boundaries, respectively, under a given precursor amplification factor P. , This is the boundary reference value; , , Let be the boundary slope coefficient, and satisfy . < < The advantage of this approach is that as the pre-monitor amplification factor increases, the threshold required for the abnormal response value to reach the same risk level can be appropriately reduced, thus reflecting the "pre-monitor amplification effect." That is, even if the current abnormal amplitude has not yet reached a high level, as long as it shows a continuous amplification trend, it can enter a higher warning level in advance.

[0120] Subsequently, based on the positional relationship between the current two-dimensional judgment point and the threshold boundaries at each level, specific criterion items in the early warning criterion set A are generated. The early warning criterion set A includes at least: low-level early warning criteria, intermediate-level early warning criteria, and high-level early warning criteria. Specifically, it can be defined as: When the abnormal response value after bounded operating condition modulation is less than the first warning threshold boundary, the normal condition criterion is met; when the abnormal response value after bounded operating condition modulation is not less than the first warning threshold boundary and is less than the second warning threshold boundary, the low-level warning criterion is met; when the abnormal response value after bounded operating condition modulation is not less than the second warning threshold boundary and is less than the third warning threshold boundary, the intermediate-level warning criterion is met; when the abnormal response value after bounded operating condition modulation is not less than the third warning threshold boundary, the high-level warning criterion is met.

[0121] The S600 inputs A into the remote monitoring system and performs coupled corrections based on the robot's task scheduling status, outputting graded early warning results and corresponding adaptive intervention strategy instructions.

[0122] After constructing the early warning criterion set A, this step further transforms the aforementioned anomaly identification results into executable early warning outputs for the actual operation scenarios of the forklift mobile robot. Since the impact of battery anomaly risks on the overall vehicle's operational safety and task continuity varies under different task loads, job priorities, and remaining job paths, this step does not directly output fixed early warning conclusions based on the early warning criterion set A. Instead, it couples and corrects the early warning criterion set A with the robot's task scheduling state to obtain tiered early warning results and corresponding adaptive intervention strategy instructions.

[0123] In specific implementation, the system first receives the early warning criterion set A corresponding to the current time point. This set includes at least the early warning level criterion satisfied at the current time point, the corresponding abnormal response value after operational condition modulation, the precursor amplification coefficient, and the duration for which the early warning level criterion is consistently met. Subsequently, the robot task scheduling status corresponding to the current time point is read. This status includes at least the following: current task type, current task priority, current loading status, current lifting status, estimated remaining operation time, estimated travel distance from the current location to the charging point, estimated travel distance from the current location to the safe unloading point, and whether the current task is allowed to be interrupted. Based on this information, it can be determined whether further development of the current battery anomaly will amplify its impact on heavy-load lifting processes, narrow aisle passage processes, fork high-position holding processes, or emergency delivery tasks.

[0124] To enable the robot's task scheduling state to participate in subsequent correction calculations, the robot's task scheduling state is first quantitatively expressed. Specifically, the current task type is divided into driving and transfer tasks, picking and lifting tasks, unloading and lowering tasks, empty return to warehouse tasks, and charging and docking tasks, and different task risk weights are assigned to each; the current task priority is divided into low priority, medium priority, and high priority, and each is assigned an increasing priority weight; the current loading state is divided into unloaded state and loaded state, and the current lifting state is divided into non-lifting state and lifting state; the estimated remaining operation time, the estimated travel distance from the current position to the charging point, and the estimated travel distance from the current position to the safe unloading point are normalized to between 0 and 1; whether the current task can be interrupted is converted into an interruptible flag value and an interruptible flag value. After completing the above processing, a task scheduling state feature sequence is formed.

[0125] Based on this, task coupling correction is performed on the early warning criterion set A. Specifically, a basic risk value is first generated according to the original early warning level corresponding to the early warning criterion set A. Preferably, the basic risk value corresponding to the low-level early warning criterion can be set to 0.3 to 0.5, the basic risk value corresponding to the medium-level early warning criterion can be set to 0.5 to 0.7, and the basic risk value corresponding to the high-level early warning criterion can be set to 0.7 to 1.0. Subsequently, the task risk weight, priority weight, cargo status identifier value, lifting status identifier value, normalized value of the estimated remaining operation time, normalized value of the estimated driving distance to the charging point, normalized value of the estimated driving distance to the safe unloading point, and interruption permission identifier value in the task scheduling state feature sequence are weighted and combined to obtain the task coupling correction value. The calculation principle of the task coupling correction value is: a positive gain is given to the state that will amplify the abnormal consequences, and a negative gain is given to the state that is conducive to quick escape or quick end of risky actions. For example, when the vehicle is under load and in a lifted state, the anomaly has a greater impact on the stability of the whole vehicle, and the corresponding correction value is increased; when the vehicle is returning to the depot empty and is close to the charging point, even if there is an anomaly, it is easier to perform conservative handling, and the corresponding correction value is reduced; when the task is a high-priority task and the task cannot be interrupted, it is necessary to balance safety and task continuity. The correction value is neither simply increased to the highest level nor directly ignored. Instead, it is determined by combining the remaining working time and the current position.

[0126] Preferably, the comprehensive risk value after task coupling correction can be determined as follows: the basic risk value and the task coupling correction value are added or weighted and then boundedly compressed to keep the comprehensive risk value between 0 and 1. When the comprehensive risk value increases significantly with task coupling correction, it indicates that the battery anomaly of the same level has a higher realistic risk in the current operating scenario; when the comprehensive risk value decreases, it indicates that there is still a large room for mitigation under the current task state.

[0127] After obtaining the comprehensive risk value, a graded early warning result is output. Preferably, the graded early warning result can be divided into four levels: Normal Attention Level, Restricted Operation Level, Priority Handling Level, and Immediate Protection Level. Specifically, when the comprehensive risk value is lower than the first grade threshold, the Normal Attention Level is output; when the comprehensive risk value is not lower than the first grade threshold but lower than the second grade threshold, the Restricted Operation Level is output; when the comprehensive risk value is not lower than the second grade threshold but lower than the third grade threshold, the Priority Handling Level is output; and when the comprehensive risk value is not lower than the third grade threshold, the Immediate Protection Level is output. The first, second, and third grade thresholds can be set based on historical operational data statistics, for example, 0.35 to 0.45, 0.55 to 0.70, and 0.80 to 0.90, respectively. To reduce level fluctuations caused by instantaneous volatility, the same graded result can be required to remain continuously for at least two to three judgment cycles before being officially output.

[0128] After outputting the graded early warning results, adaptive intervention strategy instructions corresponding to each grade are further generated. These adaptive intervention strategy instructions are not fixed and unique actions, but are selected based on both the graded early warning results and the task scheduling status. Specifically, when the normal concern level is output, a monitoring enhancement instruction is generated. This instruction is used to increase the subsequent sampling frequency, shorten the analysis cycle, and record the current abnormal evolution process. When the restricted operation level is output, a performance constraint instruction is generated. This instruction is used to limit the maximum discharge current, limit the maximum travel speed, limit the lifting speed, or prohibit continuous high-frequency start-stop. When the priority handling level is output, a task replanning instruction and a directional withdrawal instruction are generated. The task replanning instruction is used to suspend the allocation of new high-load tasks, prioritize empty return to the warehouse, nearby loading, or transition to low-intensity operations. The directional withdrawal instruction is used to control the robot to prioritize heading to a safe unloading point or the nearest charging point. When the immediate protection level is output, a protection shutdown instruction or an emergency load reduction instruction is generated. The protection shutdown instruction is used to immediately stop high-risk actions under safe conditions. The emergency load reduction instruction is used to first complete the fork descent, release the lifting load, and maintain a low-power standby state.

[0129] To ensure the scenario adaptability of the adaptive intervention strategy commands, this invention further specifies the selection order of intervention commands. If the current state is loaded and lifted, a fork descent command is generated first, followed by a low-speed retreat command, and finally a stop or retraction command, to avoid secondary risks caused by direct interruption of action under high load conditions. If the current state is unloaded and close to the charging point, a directional recharging command is generated first. If the current task cannot be interrupted but the overall risk value is at the priority level, a command to complete the current minimum closed-loop operation with limited power is generated first, and the retraction process is immediately initiated after the minimum closed-loop operation is completed. If the current overall risk value reaches the immediate protection level, regardless of the task priority, safety protection is the highest constraint, and lifting and high-load actions are released first.

[0130] Therefore, this step, by coupling and correcting the early warning criterion set A with the robot task scheduling state, not only outputs a graded early warning result that matches the current real operation scenario, but also generates targeted adaptive intervention strategy instructions. This extends the remote monitoring and early warning method of the present invention from simple anomaly identification to an executable handling scheme for battery risks of forklift mobile robots, which can prioritize the control of safety risks caused by anomaly expansion while ensuring the continuity of operation.

[0131] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A remote monitoring and early warning method for a battery system of a forklift mobile robot, characterized in that: include: S100 collects the operating data of the battery system under incomplete charge-discharge cycle conditions and constructs a composite dataset D containing voltage recovery curves, load change response curves, and static decay trajectories. S200 performs cross-condition alignment processing on D, extracts voltage rebound hysteresis features and temperature-current phase shift features at different time scales, and forms a non-synchronous feature set F. S300, based on F, construct an implicit state mapping model H of internal polarization-diffusion coupling of the battery, and obtain the equivalent characterization parameter set Q of the unobservable state quantities; S400, Q is mapped to the historical health evolution space to generate a deviation trajectory, and the deviation trajectory is self-verified by constructing a reverse consistency verification function to obtain the credible anomaly factor C; S500 generates a set of early warning criteria A with a precursor amplification effect based on the nonlinear correlation between C and real-time operating condition triggering conditions. The S600 inputs A into the remote monitoring system and performs coupled corrections based on the robot's task scheduling status, outputting graded early warning results and corresponding adaptive intervention strategy instructions.

2. The remote monitoring and early warning method for a forklift mobile robot battery system according to claim 1, characterized in that: The process of constructing the composite dataset D includes: S110 synchronously collects voltage, current and temperature data at a preset sampling frequency during battery system operation, and marks the trigger time of load change to form the original time-series data stream; S120, based on the load change trigger time, segments the original time-series data stream to obtain the corresponding load change response curve, and simultaneously extracts the voltage recovery range after the load is removed. S130 performs curve fitting on the voltage recovery range, extracts characteristic parameters representing the voltage rebound rate, and generates a static decay trajectory by combining the voltage decay data of the static phase. S140, the load mutation response curve, voltage recovery curve and static attenuation trajectory are aligned and fused according to a unified time reference to construct the composite dataset D.

3. The remote monitoring and early warning method for a forklift mobile robot battery system according to claim 1, characterized in that: The extraction of voltage rebound hysteresis features and temperature-current phase shift features at different time scales includes: S210, group the various types of data in the composite dataset D according to the trigger time of load change, and perform normalization processing on the time series under different working conditions based on a unified time benchmark to form a comparable standard time series dataset. S220, Based on the standard time series dataset, a multi-time-scale sliding window is constructed, and the voltage recovery curves in each window are differentially calculated to obtain a sequence characterizing the voltage change rate, and the voltage rebound hysteresis characteristic parameters are extracted accordingly. S230, while extracting the voltage rebound hysteresis characteristic parameters, phase matching processing is performed on the temperature data and current data within the corresponding time period to calculate the time offset between temperature change and current change, and the temperature-current phase offset characteristics are obtained.

4. The remote monitoring and early warning method for a forklift mobile robot battery system according to claim 3, characterized in that: After analyzing the extracted voltage rebound hysteresis features, a voltage rebound hysteresis anomaly index is generated. After analyzing the extracted temperature-current phase shift features, a temperature-current phase shift anomaly index is generated. The voltage rebound hysteresis anomaly index and the temperature-current phase shift anomaly index are combined to form an asynchronous feature set F.

5. The remote monitoring and early warning method for a forklift mobile robot battery system according to claim 1, characterized in that: The implicit state mapping model H based on F, which constructs the internal polarization-diffusion coupling of the battery, includes: S310, jointly encodes the voltage rebound hysteresis anomaly index and the temperature-current phase shift anomaly index in the asynchronous feature set F, constructs a multi-dimensional feature vector sequence, and forms the input feature matrix in time order; S320, Based on the input feature matrix, a state mapping function containing polarization response term and diffusion hysteresis term is constructed, and the function parameters are solved by the minimum error iteration method to obtain an implicit state mapping model H describing the feature evolution relationship; S330, Apply the implicit state mapping model H to the current input feature matrix to calculate the corresponding implicit state response sequence, and extract the intermediate state quantities that reflect the polarization intensity and diffusion hindrance. S340 performs normalization and weighted fusion processing on intermediate state variables to generate an equivalent representation parameter set Q for unobservable state variables.

6. The remote monitoring and early warning method for a forklift mobile robot battery system according to claim 1, characterized in that: The process of mapping Q to the historical health evolution space to generate deviation trajectories includes: S410, the polarization equivalent characterization parameters, diffusion equivalent characterization parameters and coupling equivalent characterization parameters in the equivalent characterization parameter set Q are input into the pre-constructed historical health evolution space in chronological order, and the corresponding reference health trajectory points are determined by the minimum distance matching method to form the current state mapping sequence; S420: Based on the difference between the current state mapping sequence and the corresponding reference health trajectory point, calculate the offset of each time point and connect them in chronological order to form the deviation trajectory.

7. The remote monitoring and early warning method for a forklift mobile robot battery system according to claim 6, characterized in that: Methods for obtaining the credible anomaly factor C include: S430, based on the deviation trajectory, constructs a reverse consistency verification function, and obtains the trajectory consistency evaluation value by back-projecting the deviation trajectory to the historical health evolution space and calculating the reconstruction error; S440 generates a reliable anomaly factor C based on the combination relationship between trajectory consistency evaluation value and offset.

8. The remote monitoring and early warning method for a forklift mobile robot battery system according to claim 1, characterized in that: The set of early warning criteria A with precursor amplification effect includes: S510 collects the real-time operating condition triggering conditions corresponding to the current time point, and performs quantization and encoding processing on the real-time operating condition triggering conditions to form an operating condition feature vector. S520 performs nonlinear coupling calculations between the operating condition feature vector and the reliable anomaly factor C to construct a correlation function containing amplification and suppression factors, and obtains the abnormal response value after operating condition modulation. S530, calculate the precursor amplification factor based on the changing trend of abnormal response values ​​in a continuous time series; S540 combines and maps the abnormal response value with the precursor amplification coefficient to generate a set of early warning criteria A containing multiple judgment thresholds.

Images