Wood chips are extensively utilised as raw material for the pulp and bio-fuel industry, and advanced material analyses may improve the processes in utilizing these products. Electrical impedance spectroscopy (EIS) combined with machine learning was used in order to analyse heartwood content of pine chips and bark content of birch chips. A novel electrode system integrated in a sampling container was developed for the testing using frequency range 42 Hz–5 MHz. Three electrode pairs were used to measure the samples in x-, y- and z-direction. Three machine learning methods were used: K-nearest neighbor (KNN), decision tree (DT) and support vector machines (SVM). The heartwood content of pine chips and bark content of birch chips were classified with an accuracy of 91% using EIS from pure materials combined with a k-nearest neighbour classifier. When using mixed materials and multiple classes, 73% correct classification for pine heartwood content (four groups) and 64% for birch bark content (five groups) were achieved.

Air-coupled ultrasound was used for assessing natural defects in wood boards by through-transmission scanning measurements. Gas matrix piezoelectric (GMP) and ferroelectret (FE) transducers were studied. The study also included tests with additional bias voltage with the ferroelectret receivers. Signal analyses, analyses of the measurement dynamics and statistical analyses of the signal parameters were conducted. After the measurement series, the samples were cut from the measurement regions and the defects were analyzed visually from the cross sections. The ultrasound responses were compared with the results of the visual examination of the cross sections. With the additional bias voltage, the ferroelectret measurement showed increased signal-to-noise ratio, which is especially important for air-coupled measurement of high-attenuation materials like wood. When comparing the defect response of GMP and FE sensors, it was found that FE sensors had more sensitive dynamic range, resulting from better s/n ratio and short response pulse. Classification test was made to test the possibility of detecting defects in sound wood. Machine learning methods including decision trees, k-nearest neighbor and support vector machine were used. The classification accuracy varied between 72 and 77% in the tests. All the tested machine learning methods could be used efficiently for the classification.

The combined method is based on using an electrical method for moisture gradient monitoring and an acoustic emission method for detection of microcracking. In the method, electrodes are used to create an electric field in drying wood and to measure the electric complex spectrum using the impedance spectroscopy method, while at the same time measuring acoustic emissions from drying wood. The result can be used for calculating the parameters required for determining the stress state prevailing in the wood. The spectrum is affected by the properties of the wood, such as moisture content, moisture gradient, temperature, density and structure. When the electric complex spectrum and acoustic emission response are determined at the same time, it is possible to estimate both the main reason for the stresses (moisture gradient) and the outcome (microcracking and macrocracking). The results may be used to control drying to achieve wood products of superior quality. The method was studied and tested both in the laboratory and in industrial kilns.

In circular economy, the use of wood chips from recycled and waste wood with varying composition has increased. Thus, the challenges with impurities have increased and it would be advantageous to be able to detect, identify and quantify the impurities. The goal of the study was to find out if wood chips with impurities of aluminum, recycled plastic or paper can be analysed using electrical impedance spectroscopy (EIS) and machine learning (ML). An EIS-measurement prototype using frequency range 1–500 kHz was developed for biomass samples. Five electrodes were used to measure the samples in two main directions. Multiple and Gaussian process regression (MR, GPR) were used for regression analyses and k-nearest neighbor (KNN), decision tree (DT), and support vector machines (SVM) for classification. In the study, frozen, unfrozen and partly unfrozen samples were used. Moisture content (MC) variation was from oven-dry to 68% (wet-basis). Recycled plastic and shredded aluminium could be classified with 95% correct classification rate (CCR), for shredded paper the CCR was 75%. The amounts of impurities (volumetric content of shredded aluminium) could be determined with 2% accuracy (RMSE) using EIS, volume weight and GPR. An interesting finding was that the classification was more accurate for frozen samples than for the unfrozen samples.